# Full Text: A Living Meta-Analysis Architecture for Active Inference: Assertion Extraction, Nanopublications, and Hypothesis Scoring

> Extracted from `act_inf_metaanalysis_v2_04-30-2026.pdf`

---

## Page 1

A Living Meta-Analysis Architecture for Active Inference
Assertion Extraction, Nanopublications, and Hypothesis Scoring
Daniel Friedman
Active Inference Institute
daniel@activeinference.institute
ORCID: 0000-0001-6232-9096
and Joel Dietz
Massachusetts Institute of Technology (MIT)
California Institute for Machine Consciousness (CIMC)
jdietz@mit.edu
ORCID: 0000-0002-9456-2691
DOI: 10.5281/zenodo.19461934
April 30, 2026
Contents
1
Abstract
5
2
Introduction: Evidence Gaps in a Rapidly Expanding Field
6
2.1
The Free Energy Principle and Active Inference Framework . . . . . . . . . . . . . . . . . . . . . . . .
6
2.2
Challenges Posed by Rapid Literature Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.3
Related Work and Prior Meta-Analyses
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.4
Synergizing Knowledge Graphs and LLMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.5
This Study: Approach and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.6
Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.7
Scope and Delimitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.8
Principal Contributions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
3
Methodology: Pipeline Design and Formal Definitions
9
3.1
Pipeline Overview
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.2
Reproducible Build Infrastructure
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.3
Stage 1: Multi-Source Literature Retrieval and Deduplication . . . . . . . . . . . . . . . . . . . . . . .
10
3.3.1
Canonical Identifier Deduplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.3.2
Relevance Filtering and Curation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.4
LLM-Based Assertion Extraction: Prompt Design, Error Taxonomy, and Validation
. . . . . . . . . .
11
3.4.1
Relationship to Prior Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
3.4.2
The Eight Tracked Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
3.4.3
Prompt Engineering and Schema Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.4.4
Failure Modes and Error Recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.4.5
Validation Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
3.4.6
From Assertions to Nanopublications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
3.5
Stage 2: Bibliometric Analysis
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.5.1
Subfield Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.5.2
Temporal Metrics and Growth-Rate Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.5.3
Text Analytics
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.5.4
Citation Network Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.6
Stage 3: Nanopublication-Based Knowledge Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
3.6.1
LLM-Based Assertion Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15

## Page 2

3.6.2
Nanopublication Schema and RDF Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
3.6.3
Knowledge Graph Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
3.6.4
Citation-Weighted Hypothesis Scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
3.6.5
Tally-Based Evidence Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
3.7
Stages 4–5: Visualization, Variable Injection, and Reproducibility
. . . . . . . . . . . . . . . . . . . .
17
3.7.1
Stage 4: Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
3.7.2
Stage 5: Manuscript Variable Injection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
3.7.3
Reproducibility and Test-Driven Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
4
Results
18
4.1
Hypothesis Evidence Landscape and Temporal Dynamics
. . . . . . . . . . . . . . . . . . . . . . . . .
18
4.1.1
Interpretation of Evidence Profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
4.1.2
Temporal Dynamics of Evidence Accumulation . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
4.1.3
Assertion Composition and Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
4.1.4
Limitations of the Current Scoring Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
4.1.5
Methodological Validation and LLM Calibration
. . . . . . . . . . . . . . . . . . . . . . . . . .
21
4.2
Field Overview: Disciplinary Structure and Growth Dynamics
. . . . . . . . . . . . . . . . . . . . . .
22
4.2.1
Corpus-Level Summary
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
4.2.2
Domain Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
4.2.3
Cross-Domain Comparison
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
4.3
Domain Analyses: Growth Trajectories and Open Problems . . . . . . . . . . . . . . . . . . . . . . . .
26
4.3.1
Domain A: Core Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
4.3.2
Domain B: Tools & Translation Methods
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
4.3.3
Domain C: Application Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
4.3.4
Comparative Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
4.3.5
Text Analytics: Topic Modeling, Vocabulary Structure, and Document Embeddings
. . . . . .
29
4.3.6
Topic Modeling: Latent Structure
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
4.3.7
Vocabulary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
4.3.8
Document Embedding Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
4.3.9
Domain Semantic Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
4.3.10 Term Co-occurrence Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
4.4
Citation Network Topology
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.4.1
Network Density and Degree Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.4.2
Connected Components and Citation Isolation
. . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.4.3
Network Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
5
Conclusion: Evidence Landscape, Methodological Limitations, and Research Agenda
36
5.1
Summary
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
5.2
Constraints and Methodological Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
5.2.1
Keyword Classifier Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
5.2.2
Citation Network Coverage Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
5.2.3
Corpus Biases, Citation Dynamics, and Linguistic Framing
. . . . . . . . . . . . . . . . . . . .
36
5.2.4
LLM Extraction Fidelity, Domain Drift, and Robustness . . . . . . . . . . . . . . . . . . . . . .
36
5.3
Research Agenda: Four Priority Next Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
5.3.1
Next Step 1 — Expand the Scope of Referenced Data . . . . . . . . . . . . . . . . . . . . . . .
37
5.3.2
Next Step 2 — Extract and Verify Evidence Supporting Claims in Each Paper . . . . . . . . .
37
5.3.3
Next Step 3 — Tie Hypotheses to Real-World Outcomes . . . . . . . . . . . . . . . . . . . . . .
38
5.3.4
Next Step 4 — Formal Evaluation Rubric for Pipeline Quality
. . . . . . . . . . . . . . . . . .
38
5.4
Future Directions: Beyond Tally-Based Evidence Aggregation . . . . . . . . . . . . . . . . . . . . . . .
39
5.4.1
Hierarchical Bayesian Hypothesis Scoring
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
5.4.2
Causal Evidence Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
5.4.3
Evidential Diversity and Source Weighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
5.4.4
Agentic LLM Extractors and Domain Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . .
39
5.5
Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
5.6
Broader Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
6
Discussion: Implications and Community Recommendations
41
6.1
Relationship to Prior Development Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
2

## Page 3

6.2
Tactical and Strategic Priorities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.2.1
Adopt Rigorous Reporting Metadata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.2.2
Explore Open Knowledge Graph Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.2.3
Standardize the Ontological Lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.3
Empirical and Theoretical Imperatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.3.1
Architect Unified Performance Benchmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.3.2
Prioritize Empirical Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.4
Living Review Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.4.1
Agentic Workspaces and MCP Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.4.2
The Discovery Engine and Future Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.5
Open Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.6
Pipeline as a Community Instrument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
6.7
Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
7
Appendix: Tooling and Infrastructure
44
7.1
LLM-Based Assertion Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
7.2
Software Ecosystem
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
7.2.1
General-Purpose Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
7.2.2
Deep Active Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
7.2.3
Predictive Coding and Neural Generative Coding . . . . . . . . . . . . . . . . . . . . . . . . . .
45
7.2.4
Benchmarking Progress
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
7.2.5
Comprehensive Open-Source Tool Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
7.2.6
Comparative Feature Matrix
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
7.3
Knowledge Graph Infrastructure
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
7.4
Multi-Level Quality Assurance
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
7.4.1
Assertion-Level Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
7.4.2
Graph-Level Consistency Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
7.4.3
Score-Level Unit Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
7.4.4
Pipeline-Level Test Coverage
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
7.4.5
Quality Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
8
Appendix: Mathematical and Algorithmic Details
50
8.1
Citation-Weighted Hypothesis Scoring Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
8.2
Non-negative Matrix Factorization (NMF) for Topic Modeling
. . . . . . . . . . . . . . . . . . . . . .
50
8.3
Field Growth-Rate Estimation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
8.4
Advanced Visualization Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
8.4.1
PCA of TF-IDF Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
8.4.2
Hierarchical Clustering Dendrogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
8.4.3
Term Heatmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
8.4.4
Term Co-occurrence Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
8.5
Nanopublication RDF Schema
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
8.5.1
Namespace Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
8.5.2
Core Triple Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
9
Appendix: Accessibility, Cognitive Ergonomics, and Participatory Research Infrastructure
54
9.1
Cognitive Ergonomics of Knowledge Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
9.1.1
Action–Intention UX and Active Inference Design Principles . . . . . . . . . . . . . . . . . . . .
54
9.1.2
Risk-Aware and Bias-Transparent Design
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
9.2
FAIR Data and Decentralized Science
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
9.3
Participatory Research and Universal Access
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
9.4
Pipeline Accessibility Checklist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
10 Notation, Abbreviations, and Glossary
57
10.1 Mathematical Symbols and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
10.2 Abbreviations and Acronyms Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
10.3 Standard Hypothesis Definitions and Identifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
10.4 Glossary of Key Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
3

## Page 4

11 References
63
4

## Page 5

1
Abstract
No prior automated system tracks hypothesis-level evidence across the full Active Inference and Free Energy Principle
(FEP) literature. Manual synthesis cannot keep pace with a field that has grown at a compound annual rate of 20.36%
across 2005–2026, and the FEP’s theoretical generality has invited falsifiability critiques that only hypothesis-specific
evidence profiling can address. Building on pioneering systematic manual annotation paired with ontology-based anal-
ysis at the scale of hundreds of papers, we present a computational meta-analysis framework that automates and scales
this approach. The pipeline retrieves literature from arXiv, Semantic Scholar, and OpenAlex, deduplicating 𝑁= 819
papers via a canonical identifier hierarchy (DOI > arXiv ID > Semantic Scholar ID > OpenAlex ID). It classifies papers
into a three-tier taxonomy spanning eight categories: A (Core Theory), B (Tools & Translation), and C (Application
Domains). An LLM-powered extraction system then evaluates each abstract against eight core hypotheses, producing
structured nanopublications—each encoding directionality, a confidence score, and natural-language reasoning—that
populate an RDF-compatible knowledge graph scored by a citation-weighted evidence function.
All extracted assertions are automatically generated and have not been manually validated; hypothesis scores should
be considered preliminary.
The resulting evidence landscape reveals a field where application domains (Domain C, 64.0%) collectively dominate the
corpus, with tools development (Domain B, 20.8%)—including pymdp, RxInfer.jl, and interpretable alternatives such
as Free Energy Projective Simulation—and core theory (Domain A, 15.2%) rounding out the taxonomy. Non-negative
matrix factorization identifies 5 latent topics that cross-cut the keyword taxonomy, and citation network analysis
exposes a sparse yet structured graph (2,176 intra-corpus edges out of 29,323 total outgoing references—only 7.4%
reference resolution, reflecting the corpus’s specialised scope rather than the underlying citation density of any single
paper) anchored by pronounced hub papers. Hypothesis scores cluster into three tiers: a broad consensus tier (score
> 0.83) covering five hypotheses—H7 Morphogenesis, H2 AIF Optimality, H4 Predictive Coding, H6 Clinical Utility,
and H5 Scalability; a near-consensus boundary (H8 Language AIF, score ≈+0.83); a moderate debate tier (H3
Markov Blanket Realism, ≈+0.78); and a diffuse tier (H1 FEP Universality, ≈+0.48) where a large neutral plurality
reflects the principle’s broad invocation without explicit empirical test—though absolute score magnitudes are inflated
by publication bias and linguistic asymmetry in academic writing, making relative rankings and temporal trajectories
more reliable than point estimates. By demonstrating that automated LLM-driven assertion extraction—operating
without human-validated ground truth—can generate scalable, queryable representations of scientific evidence, this
work provides a reusable architecture for living literature reviews—continuously updated knowledge graphs that track
hypothesis-level consensus across rapidly evolving fields.
Keywords: Active Inference, Free Energy Principle, meta-analysis, knowledge graph, nanopublications, bibliometrics,
hypothesis scoring, LLM extraction, computational neuroscience
5

## Page 6

2
Introduction: Evidence Gaps in a Rapidly Expanding Field
2.1
The Free Energy Principle and Active Inference Framework
The Free Energy Principle (FEP), introduced by Karl Friston, proposes that self-organizing systems maintain their
structural and functional integrity by minimizing variational free energy—an upper bound on sensory surprise [Friston
et al., 2006, Friston, 2010]. Under this principle, living systems are cast as approximate Bayesian inference engines
that build generative models of their environment and act to reduce the discrepancy between predicted and observed
states. Active Inference (AIF) extends this picture from passive perception to goal-directed behavior: agents select
actions that bring about observations consistent with their preferred states, unifying perception, learning, and decision-
making within a single variational framework [Parr et al., 2022, Friston et al., 2017]. Since its initial formulation for
sensorimotor control, AIF has been applied to navigation, visual foraging, language comprehension, social cognition,
and multi-agent coordination. Bayesian mechanics [Sakthivadivel, 2023] has further strengthened the mathematical
foundations of the FEP by grounding Markov blanket dynamics in the physics of belief-based systems, placing the
principle on a footing commensurate with established physical theories.
Importantly, the variational free energy
minimization at the core of the FEP shares deep mathematical connections with the broader family of Energy-Based
Models (EBMs) [LeCun et al., 2006]—including Helmholtz machines [Dayan et al., 1995], Boltzmann machines [Hinton,
2002], and variational autoencoders [Kingma and Welling, 2014]—all of which parameterize learning and inference
through scalar energy functions and variational bounds. This convergence motivates the inclusion of EBM-adjacent
literature in our search scope.
2.2
Challenges Posed by Rapid Literature Growth
The active inference literature has grown at a compound annual rate of 20.36% across 2005–2026, with annual output
accelerating sharply after 2013. While early research concentrated on theoretical neuroscience, the field has since
diversified across biology (C5), robotics (C2), computational psychiatry (C4), algorithm scaling (B), and formal math-
ematics (A1).
With 𝑁= 819 papers spanning 8 categories across 3 domains, no prior automated system tracks
hypothesis-level evidence across the full corpus.
This creates three interrelated challenges.
First, the balance of
evidence for core claims—such as FEP universality or the physical realism of Markov blankets—cannot be assessed
without structured, hypothesis-specific extraction at corpus scale. Second, because the relationship between mathe-
matical formalisms and empirical evidence is frequently implicit, systematic evidence synthesis demands substantial
manual effort: Knight et al. [Knight et al., 2022] required human annotators to manually code hundreds of papers.
Third, new entrants must navigate a literature weighted toward broad qualitative philosophy (A2), interspersed with
specialized applied subfields whose findings are difficult to locate without domain-specific search strategies.
Traditional narrative reviews attempt to address these challenges but are static, subjective, and quickly outdated.
Systematic reviews from evidence-based medicine offer rigorous aggregation but are structured for clinical trial data
with homogeneous outcome measures, making them poorly suited for the heterogeneous ontological and computational
claims in this literature. The expansion of predictive processing [Clark, 2013, Hohwy, 2013] and the emergence of
Bayesian mechanics [Sakthivadivel, 2023] further broaden the scope of assertions that a comprehensive meta-analysis
must reconcile. Critically, the falsifiability of the FEP itself remains contested [Colombo and Seriès, 2021]: because
free energy minimization can be reframed to accommodate any behavior post hoc, distinguishing genuine predictive
commitment from tautological redescription requires exactly the hypothesis-specific, evidence-quantified framework
we propose here.
2.3
Related Work and Prior Meta-Analyses
Several prior efforts have surveyed aspects of the Active Inference landscape. Sajid et al. [Sajid et al., 2021] compare
active inference with alternative decision-making frameworks; Da Costa et al. [Da Costa et al., 2020] synthesize the
discrete-state-space formulation; Lanillos et al. [Lanillos et al., 2021] survey robotics applications; Smith et al. [Smith
et al., 2022] provide a tutorial bridging theory and empirical data; and Millidge et al. [Millidge et al., 2021] examine
information-theoretic foundations of exploration behavior. Ramstead et al. [Ramstead et al., 2018] extend the FEP
to questions of biological self-organization, while Pezzulo et al. [Pezzulo et al., 2015] connect active inference to
homeostatic regulation. Millidge [Millidge, 2024] provides a practitioner’s retrospective confirming that AIF’s strongest
demonstrated results arise from novel discrete generative models, while scalability relative to deep reinforcement
learning remains the field’s central open challenge.
Parallel to these synthesis efforts, Sanjeev V. Namjoshi’s 2026 textbook, Fundamentals of Active Inference [Namjoshi,
2026b], provides a comprehensive, computationally explicit foundation for the field designed for engineers. In conjunc-
6

## Page 7

tion with this text, Namjoshi developed the aif-fep-db repository [Namjoshi, 2026a]—an open-source, dynamically
updated database of scraped and tagged publications covering active inference, the free energy principle, and predic-
tive processing. While aif-fep-db curates and categorizes the literature to facilitate reproducible systematic reviews
and interactive Dash-based exploration, it functions primarily as a modular bibliographic foundation rather than an
automated hypothesis evaluation engine.
Closest to our work, Knight, Cordes, and Friedman [Knight et al., 2022] conducted a systematic literature analysis
of publications using the terms “Free Energy Principle” or “Active Inference,” with an emphasis on works by Karl
J. Friston. Their analysis—maintained by the Active Inference Institute—combined manual annotation of structural,
visual, and mathematical features with automated analyses using the Active Inference Ontology at the scale of thou-
sands of citations and hundreds of annotated papers. That study identified six development directions—including
broader scope, richer annotation, and transferable approaches—and represents an important precursor to automated
meta-analysis of this field.
These prior works differ from the present study along four dimensions. First, scale: narrative reviews cover tens to
low hundreds of papers; our pipeline processes 𝑁= 819. Second, structure: prior reviews produce prose summaries
rather than machine-queryable knowledge graphs with typed relationships. Third, temporal tracking: no prior
system computes how evidence for specific hypotheses evolves year over year. Fourth, automation: the systematic
analysis of Knight et al. [Knight et al., 2022] pioneered quantitative literature analysis but relied on manual annotation,
limiting update frequency. Our framework advances this line of work by (1) fully automating assertion extraction via
LLM-based hypothesis scoring, (2) constructing a structured, RDF-compatible knowledge graph scored by citation-
weighted evidence, and (3) tracking how evidence for core claims evolves over time through temporal trend analysis.
2.4
Synergizing Knowledge Graphs and LLMs
Broadening this synthesis, recent systematic literature initiatives underscore a powerful reciprocal synergy between
Large Language Models (LLMs) and Knowledge Graphs: LLMs parse unstructured text to rapidly extract semantic
claims, efficiently populating the structured, queryable architecture of the graph [Quevedo Tumailli et al., 2025, Li
et al., 2024]. We adopt the nanopublication [Groth et al., 2010]—a minimal, machine-readable unit of scientific evidence
comprising a core assertion bound to explicit provenance metadata—as the fundamental serialization format for this
extracted knowledge.
2.5
This Study: Approach and Overview
This paper presents a computational meta-analysis of the Active Inference literature (𝑁= 819). Rather than relying
exclusively on bibliometric metadata or slow manual coding, we deploy a Large Language Model (LLM) to “read”
each paper’s abstract and assess its relationship to eight core hypotheses within the FEP paradigm. We serialize
these assessments as nanopublications—each encoding an assertion (“Paper X supports Hypothesis Y”) coupled with
the LLM’s natural-language reasoning and confidence score. The resulting knowledge graph aggregates these nanop-
ublications and links them to paper metadata, citation networks, subfield classifications, and hypothesis definitions.
A citation-weighted scoring formula quantifies the net evidence for or against each hypothesis, producing scores in
[−1, 1] that reflect both the direction and strength of published evidence. Importantly, this represents an open-source
introductory analysis which will be augmented and extended, and stewarded in collaborative development by the
Active Inference Institute (activeinference.org).
2.6
Research Questions
This meta-analysis addresses four primary research questions:
1. RQ1 (Field Structure): What is the disciplinary structure and growth trajectory of the Active Inference
literature, and how are papers distributed across the three domains—Core Theory (A), Tools & Translation (B),
and Application Domains (C)? We expect Domain A to dominate but anticipate measurable diversification into
applied domains.
2. RQ2 (Growth Dynamics): What are the temporal growth dynamics of the field, and which subfields are
experiencing the most rapid expansion? Prior reviews suggest accelerating growth post-2013; we quantify this
trajectory and identify which application domains drive it.
3. RQ3 (Hypothesis Evidence): What is the current balance of evidence for and against the eight standard
hypotheses, and how has this balance evolved over time? We expect well-established hypotheses (H4 Predictive
7

## Page 8

Coding) to show consensus while philosophically contested claims (H3 Markov Blanket Realism) show mixed
evidence. (See hypothesis dashboard and assertion figures in the hypothesis results.)
4. RQ4 (Tooling Readiness): What is the state of software tooling and infrastructure for Active Inference
research, and what gaps remain? We survey available implementations to identify whether the ecosystem is
fragmented or converging.
2.7
Scope and Delimitations
This study focuses on the English-language peer-reviewed and preprint literature retrievable from arXiv, Semantic
Scholar, and OpenAlex. Our search scope begins at 2005—chosen to capture Energy-Based Model and variational
Bayesian antecedents (Helmholtz machines, VAEs, early Bayesian brain formulations [Dayan et al., 1995, LeCun et al.,
2006]) that share deep mathematical foundations with variational free energy minimization and predated the Free
Energy Principle label introduced in 2006 [Friston et al., 2006]. The scope includes both the core Active Inference
and Free Energy Principle literature and adjacent Energy-Based Model research where it intersects with variational
inference or generative modeling—capturing the growing convergence between these traditions. We do not include book
chapters or monographs not indexed by these sources, software documentation, or non-English publications. Domain
classification uses keyword matching (200+ indicators across 8 categories) rather than expert annotation—a deliberate
trade-off favoring reproducibility over precision, whose consequences we quantify in the field overview. Hypothesis
scoring relies on LLM-extracted assertions operating on abstracts only; claims embedded in method sections, discussion
paragraphs, or supplementary materials are not captured, and the fraction of relevant evidence residing in these sections
is unknown. The fidelity and limitations of abstract-only extraction are examined in the extraction pipeline section.
The hypothesis definitions and domain taxonomy are informed by, but not identical to, the Active Inference Ontology
used by Knight et al. [Knight et al., 2022]; future alignment would enable direct comparison with that earlier analysis.
2.8
Principal Contributions
This work makes five contributions:
1. A multi-source retrieval and deduplication pipeline for Active Inference literature, using a canonical
identifier hierarchy across three academic databases.
2. A nanopublication-based knowledge graph schema encoding directed, confidence-scored assertions about
eight core hypotheses with full provenance tracking.
3. A quantitative field overview characterizing the growth, domain distribution (A/B/C taxonomy), citation
topology, and latent topic structure of the Active Inference literature, with specific attention to how recent
benchmark results (detailed in the domain analyses) are reshaping the scalability and application landscape.
4. An LLM-based hypothesis scoring dashboard that produces differentiated evidence profiles with temporal
trend visualization.
5. A tooling assessment of the software ecosystem supporting Active Inference research, including the imple-
mented extraction pipeline, existing software (pymdp, SPM, RxInfer.jl), and knowledge graph infrastructure.
The remainder of this paper is organized as follows.
The methodology section describes the five-stage pipeline—
the central contribution enabling reproducible, automated evidence synthesis—with separate treatments of literature
retrieval, LLM-based assertion extraction, bibliometric analysis, the nanopublication-based knowledge graph, and
visualization and variable injection. The hypothesis evidence landscape presents quantitative scoring results (RQ3),
followed by the field overview with domain-level analysis (RQ1, RQ2), detailed domain analyses, text analytics,
and citation network topology. The conclusion addresses limitations and future directions; the discussion provides
community recommendations and open questions. Appendix 8 collects mathematical and algorithmic details; Appendix
7 surveys the tooling landscape (RQ4).
8

## Page 9

3
Methodology: Pipeline Design and Formal Definitions
This section describes the five-stage computational meta-analysis pipeline. Each stage corresponds to a tested, in-
dependently executable script that reads upstream outputs and produces structured artifacts. The pipeline extends
the systematic literature analysis approach of Knight et al. [Knight et al., 2022]—which combined manual annotation
with ontology-based automated analysis—by substituting manual coding with fully automated, LLM-driven assertion
extraction and citation-weighted hypothesis scoring. All code, configuration files, and reproducibility instructions—
including a Dockerized execution environment to guarantee dependency isolation—are publicly available in the project
repository (https://github.com/ActiveInferenceInstitute/act_inf_metaanalysis).
3.1
Pipeline Overview
The five-stage pipeline is summarized in Table 1.
Table 1: Five-stage computational meta-analysis pipeline. Each stage corresponds to an independently executable
script that reads upstream outputs and produces structured artifacts. Cross-references link to detailed methodology
sections.
Stage
Script
Primary Input
Primary Output
1
01_literature_search.py
API queries
corpus.jsonl
2
02_meta_analysis_pipeline.py
corpus.jsonl
Classification, temporal, TF-IDF, NMF, citation network J
3
03_build_knowledge_graph.py
corpus.jsonl
nanopublications.jsonl, nanopublications.trig, scor
4
04_generate_figures.py
All Stage 2–3 JSONs
16 publication-ready PNGs
5
05_inject_variables.py
All output JSONs
Rendered manuscript Markdown
Scripts act as thin orchestrators that import methods from tested library modules and handle file I/O. All computation
resides in the src/ packages; no analysis logic is embedded in scripts. End-to-end pipeline execution completes in
under one hour on commodity hardware (excluding LLM extraction, which depends on model size and inference
backend); all stochastic components use fixed random seeds for deterministic reproduction.
3.2
Reproducible Build Infrastructure
The five-stage analysis pipeline described above is embedded within template/ [Friedman, 2026a,b], an open-source
Infrastructure-as-Code system for computational research that turns a full research compendium—code, data, tests,
manuscript, and provenance—into a single, version-controlled, deterministically buildable repository with an enforced,
test-gated publication pipeline. template/ applies the principle of Infrastructure as Code to the research lifecycle, mak-
ing the manuscript, test suite, and provenance chain independently verifiable. The system operationalizes FAIR4RS
principles [Wilkinson et al., 2016] and supply-chain-style provenance for manuscripts, targeting structural causes of the
reproducibility crisis: fragmented workflows across LaTeX, notebooks, and ad-hoc scripts, lack of end-to-end testing,
and no binding between code, data, figures, and the final PDF.
The system employs a Two-Layer Architecture: a globally shared infrastructure layer (12 subpackages, approximately
150 Python modules) provides generic services—logging, rendering, validation, steganographic watermarking, report-
ing, and LLM integration—while self-contained project workspaces (including the present meta-analysis) carry their
own manuscript/, scripts/, src/, tests/, data/, and output/ directories, discovered purely by filesystem con-
vention. An eight-stage build pipeline enforces an ordered sequence from environment setup through test execution
(at least 90% coverage for project code, at least 60% for shared infrastructure), analysis execution, PDF rendering
(Pandoc to LaTeX to XeLaTeX with biber), output validation, LLM review, and executive reporting. A Zero-Mock
testing policy requires all tests to exercise real filesystem operations, real subprocess calls, and real computation—no
unittest.mock doubles—making test adequacy a publication gate rather than a best-effort guideline. Cryptographic
provenance is embedded in every PDF via SHA-256 hash manifests, PDF metadata injection, and optional QR codes
linking back to the repository. A Documentation Duality standard equips every directory with both human-readable
README.md and machine-readable AGENTS.md files, while each infrastructure module carries a SKILL.md skill descriptor
aligned with the Model Context Protocol, enabling AI agents to locate and invoke module capabilities without hal-
lucinating API signatures. The template/ framework and this meta-analysis project are available under the Apache
2.0 License at https://github.com/ActiveInferenceInstitute/act_inf_metaanalysis.
9

## Page 10

3.3
Stage 1: Multi-Source Literature Retrieval and Deduplication
We retrieve papers from three complementary academic databases to maximize coverage and enable cross-source dedu-
plication. The retrieval window begins at 2005, encompassing the period when Energy-Based Model and variational
Bayesian research [Dayan et al., 1995, LeCun et al., 2006] provided mathematical precursors to what Friston formal-
ized as the Free Energy Principle in 2006 [Friston et al., 2006]; this inclusive start captures historical lineage and
cross-disciplinary convergence that a later cutoff would exclude.
arXiv. We query the arXiv Atom API using five phrase-matched searches by default: all:"active inference",
all:"free energy principle",
all:"predictive coding" AND all:"free energy",
all:"expected free
energy", and all:"variational free energy" AND all:"inference". The all: prefix searches titles, abstracts,
and full text; phrase matching reduces contamination from unrelated physics papers that mention “free energy” in
thermodynamic contexts.
Additional Energy-Based Model queries (all:"energy-based model" AND all:"free
energy", all:"Helmholtz machine" AND all:"inference", all:"Boltzmann machine" AND all:"free energy",
all:"contrastive divergence" AND all:"generative model") are available via the arxiv_queries list in
config.yaml for researchers wishing to capture adjacent EBM literature at the intersection of energy-based
generative modeling and variational inference [LeCun et al., 2006].
Semantic Scholar. We query the Semantic Scholar Graph API [Kinney et al., 2023] with the same terms. Semantic
Scholar provides citation graphs, abstract embeddings, and links to published versions. Retry logic with exponential
backoff handles rate limiting.
OpenAlex. We query OpenAlex [Priem et al., 2022] to capture journal-published work that may not appear on
arXiv, including clinical studies and neuroscience experiments in domain-specific venues. The referenced_works
field populates citation links for each paper.
3.3.1
Canonical Identifier Deduplication
After retrieval, papers are assigned a canonical identifier using the priority scheme: DOI > arXiv ID > Semantic
Scholar ID > OpenAlex ID > title hash. When the same paper appears in multiple sources, the record with the
highest metadata completeness is retained. For each incoming paper, the two records are compared on metadata
completeness—defined as the count of non-empty optional attributes across the full Paper record (abstract, DOI,
arXiv ID, Semantic Scholar ID, OpenAlex ID, venue, citation count, references, publication date, PDF URL, open-
access flag, and author list). The pipeline retains the richer record; in the event of a tie, the incumbent is preserved.
This “merge-on-add” strategy aggregates the richest available metadata without requiring an expensive downstream
reconciliation pass. Deduplication produces 𝑁= 819 unique papers spanning 2005–2026.
3.3.2
Relevance Filtering and Curation
After deduplication, a relevance filter removes papers whose titles and abstracts lack any core Active Inference
terminology (e.g., active inference,''free energy principle,’ ’ “variational free energy’ ’), eliminating off-topic results
introduced by broad keyword overlap across heterogeneous databases. We acknowledge that this retrieval strategy
yields limited bibliographic depth, functioning as a representative snapshot rather than an exhaustive census of the
literature.
We emphasize that this process relies on keyword search strategies across divergent APIs. In any complex research
field, there is no single optimal word or threshold for definitive inclusion or exclusion. Different information sources
and repositories yield differing schemas and representations, introducing both false positives (e.g., machine learning
papers that mention “free energy” in a purely thermodynamic context, or bioinformatics tools whose names overlap
with AIF terminology) and false negatives (e.g., predictive coding studies that avoid the phrase “free energy principle”
entirely, or agent-based modeling papers that implement functionally equivalent algorithms under different labels). The
keyword lists in config.yaml document all search terms explicitly to enable systematic replication and refinement.
Consequently, this pipeline is not intended to produce a static, “golden” list of canonical papers. Rather, it is designed
as an open-source software package that can be modularly updated and versioned. Researchers can configure the
pipeline to operate on custom literature bibliographies curated for specific relevance criteria through time, treating
the initial query-based retrieval as a programmatic starting point rather than an absolute boundary. For example,
adding a ninth domain category (e.g., “D: Education”) requires only adding a keyword list to the subfield_keywords
section of config.yaml—no source code modification is needed.
10

## Page 11

3.4
LLM-Based Assertion Extraction: Prompt Design, Error Taxonomy, and Valida-
tion
This supplementary section documents the implementation specifics of the LLM-based assertion extraction pipeline.
3.4.1
Relationship to Prior Approaches
The closest prior effort is the systematic literature analysis of Knight, Cordes, and Friedman [Knight et al., 2022], which
used human annotators to manually code structural, visual, and mathematical features of FEP and Active Inference
publications. Their work operated at the scale of hundreds of annotated papers and employed terms from the Active
Inference Institute’s Active Inference Ontology for automated text analysis. Our pipeline replaces the manual coding
step with LLM-based assertion extraction, enabling scalable processing of the full corpus (𝑁= 819 papers) at the
cost of exchanging human-verified precision for machine-generated assessments that require post-hoc validation. This
trade-off is characteristic of the broader LLM-based scientific extraction landscape: recent benchmarking confirms that
even state-of-the-art modular extraction architectures fall short of production-level precision—particularly on tasks
requiring exhaustive retrieval and aggregation of multiple values from long documents—validating our design choice
to retain human review pathways alongside automated extraction.
Table 2: Comparison of annotation approaches: Knight et al. (2022) manual coding versus this work’s automated
LLM-based extraction pipeline. Key trade-offs between human-verified precision and machine-generated scalability
are highlighted.
Dimension
Knight et al. (2022)
This work
Scale
Hundreds of papers
819 papers
Annotation
Manual (structural/visual/math features)
Automated (LLM hypothesis assessment)
Ontology
Active Inference Ontology terms
8 standard hypotheses
Output
Annotated features + term frequencies
Nanopublications + knowledge graph
Reproducibility
Annotator-dependent
Deterministic (given model + seed)
Precision
High (human-verified)
Medium (requires validation)
3.4.1.1
Positioning in the LLM-Based Review Landscape
Our pipeline operates within a rapidly maturing
ecosystem of LLM-powered literature analysis tools. Multi-agent architectures such as LitLLM decompose the review
process into specialized sub-agents (planner, identifier, extractor, compiler), while ensemble approaches aggregate
outputs from multiple LLMs via weighted voting to improve reliability. Our work differs from these tools in three
respects: (1) we target hypothesis-level evidence scoring rather than inclusion/exclusion screening; (2) we produce
structured nanopublications rather than narrative summaries; and (3) we are only analyzing abstracts for claims. This
deliberate trade-off enables corpus-scale processing (𝑁= 819) but fundamentally misses fine-grained claims embedded
in method sections or discussion paragraphs. Full-text processing could improve extraction recall, particularly for
hypotheses with small evidence bases (H6 Clinical Utility, H7 Morphogenesis).
3.4.2
The Eight Tracked Hypotheses
Our analysis tracks the evolving evidence base for eight distinct claims within the Active Inference literature, spanning
theoretical universality to applied clinical utility:
1. H1: FEP Universality (Theoretical). The Free Energy Principle applies universally to all self-organizing
systems.
2. H2: AIF Optimality (Computational). Active Inference agents achieve optimal decision-making under
uncertainty.
3. H3: Markov Blanket Realism (Philosophical). Markov blankets correspond to real physical boundaries.
4. H4: Predictive Coding (Empirical). Cortical hierarchies minimize prediction errors via predictive coding.
5. H5: Scalability (Computational). Active Inference scales to complex, high-dimensional environments.
6. H6: Clinical Utility (Applied). Active Inference provides clinically useful models of psychiatric conditions.
7. H7: Morphogenesis (Biological). The FEP explains morphogenetic and developmental processes.
8. H8: Language AIF (Applied). Active Inference provides a viable framework for language processing.
11

## Page 12

3.4.3
Prompt Engineering and Schema Design
The structured prompt is designed to minimize parsing failures and maximize assessment quality:
1. Explicit JSON schema. The prompt specifies the exact output schema—field names, allowed direction values,
and the numeric confidence range—reducing the LLM’s tendency to generate free-form text or ad hoc structures.
2. Hypothesis definitions in-context. All eight definitions are included verbatim, ensuring the LLM assesses
relevance from the provided context rather than relying on parametric knowledge that may be stale.
3. Reasoning field. Each assessment includes a natural-language reasoning string, providing an audit trail for
human reviewers and enabling systematic analysis of error patterns.
4. Irrelevant filtering. An explicit “irrelevant” direction allows the LLM to mark hypotheses that a paper does
not address, avoiding forced spurious assessments.
3.4.3.1
Prompt Template
The extraction prompt follows a two-part structure (system + user):
SYSTEM: You are a scientific literature analyst specializing in the
Free Energy Principle and Active Inference. Assess the relevance of
the given paper to each hypothesis. Return a JSON array.
USER:
Paper: {title}
Abstract: {abstract}
Hypotheses:
H1: FEP Universality — {description}
H2: AIF Optimality — {description}
...
H8: Language AIF — {description}
For each hypothesis, return:
{
"hypothesis_id": "H1",
"direction": "supports|contradicts|neutral|irrelevant",
"confidence": 0.0-1.0,
"reasoning": "..."
}
The extraction module (src/knowledge_graph/llm_extraction.py) includes configurable retry logic with exponen-
tial backoff, JSON parsing with handling of markdown code fences and extraneous text, confidence clamping, and
validation against the hypothesis ID set. The default model is gemma3:4b on a local Ollama instance, configurable via
--llm-model and --llm-url flags.
3.4.4
Failure Modes and Error Recovery
The primary failure modes are documented below.
3.4.4.1
Over-Extraction Bias
Preliminary experiments indicated ~15–20% over-extraction. The current 𝑁=
819 corpus was processed without a validation set; error rates are not quantified for this run. This is the most common
error mode and produces false supporting evidence. Over-extraction disproportionately affects broad-scope hypotheses
(H1 FEP Universality, H2 AIF Optimality) where most papers in the corpus contain relevant terminology without
explicitly endorsing the claim. Narrower hypotheses tied to specific domains (H7 Morphogenesis, H8 Language AIF)
show lower over-extraction rates because their vocabulary is more distinctive. This systematic bias inflates support
counts for broad hypotheses, and we caution against interpreting absolute scores for H1 and H2 without accounting
for this effect.
3.4.4.2
Direction Misclassification
The LLM misclassifies a contradicting claim as supporting, or vice versa.
Rarer but more consequential, as it directly inverts the evidence signal. Most common for papers that discuss limita-
tions while ultimately endorsing a hypothesis.
12

## Page 13

3.4.4.3
Confidence Calibration Constraints
The model occasionally assigns high confidence to assessments
where the underlying evidence is ambiguous. Reliable confidence calibration remains an open problem for zero-shot
LLM applications, motivating the multi-tiered validation protocols described below.
3.4.4.4
Progressive JSON Parsing Recovery
To mitigate formatting inconsistencies, the module implements
a progressive parsing pipeline to recover malformed LLM outputs:
1. Direct parse: Attempt json.loads() on the raw response.
2. Strip code fences: Remove Markdown ```json ... ``` wrappers and retry.
3. Extract JSON array: Scan for the first [...] substring in the response text.
Papers that fail all parsing stages are logged and skipped; their count is reported at pipeline completion.
3.4.5
Validation Methodology
Validation of LLM-extracted assertions follows a three-tier protocol:
1. Validation Dataset (10%, not yet created). A ground-truth validation protocol is specified in which a
random 10% subset of the corpus will be manually annotated by human experts. Inter-rater reliability will be
calculated using Cohen’s 𝜅; the LLM-based extraction pipeline will be evaluated against this human consensus,
targeting a 𝜅> 0.70 threshold for direction accuracy (supports/contradicts/neutral/irrelevant). The formal 10%
manual annotation dataset has not yet been created; its development is a prioritized next step for this living
review architecture.
2. Boundary-case audit (conceptual design). Papers known to make contested claims (e.g., critiques of FEP
universality, Markov blanket realism debates) would be specifically checked for correct direction assignment.
This tier remains a conceptual design and has not been executed.
3. Aggregate consistency (conceptual design).
Hypothesis scores would be compared against qualitative
expectations from the literature: hypotheses known to be well-supported (e.g., H4 Predictive Coding) should
score positively; those known to be contested (e.g., H3 Markov Blanket Realism) should show lower or mixed
scores. This tier also remains a conceptual design and has not been executed.
The current extraction pipeline operates without human-validated ground truth; all reported assertions are machine-
generated and unaudited.
3.4.6
From Assertions to Nanopublications
Each validated assertion is wrapped in a nanopublication [Groth et al., 2010, Kuhn et al., 2016]—a self-contained,
machine-readable knowledge unit packaging the assertion with explicit provenance metadata. The wrapping process
assigns:
• A unique identifier (nanopub:<uuid12>) for graph-level deduplication.
• An attribution string recording the pipeline name and LLM model version.
• A UTC timestamp in ISO 8601 format, establishing temporal provenance.
Nanopublications are persisted incrementally during extraction. Every 50 papers (configurable via --checkpoint-interval),
the pipeline atomically appends newly extracted nanopublications to nanopublications.jsonl using a temporary-
file-plus-rename strategy that prevents corruption on interruption.
Deduplication operates on the composite key
(𝑝𝑎𝑝𝑒𝑟_𝑖𝑑, ℎ𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠_𝑖𝑑): when a paper is re-processed with an improved model, the newer assertion overwrites
the stale entry. This merge-on-add design enables iterative model refinement without costly full-corpus re-extraction.
After extraction completes, the full nanopublication set is additionally serialized to RDF/TriG format per the
nanopublication standard, producing four named graphs per nanopublication (Head, Assertion, Provenance, Publica-
tion Info). The TriG output is suitable for publication to the decentralized nanopublication network and archival on
data repositories such as Zenodo. The complete RDF schema is specified in the knowledge graph methodology and
Appendix 8.5.
13

## Page 14

3.5
Stage 2: Bibliometric Analysis
Stage 2 performs four complementary analyses on the deduplicated corpus. All analyses are deterministic given fixed
random seeds and operate on the same corpus.jsonl input.
3.5.1
Subfield Classification
Each paper is classified into one of eight categories organized across three domains: A – Core Theory (A1: quantita-
tive and formal mathematical theory; A2: qualitative philosophy and general FEP theory), B – Tools & Translation
(algorithms, scaling, and software development), and C – Application Domains (C1: neuroscience, C2: robotics,
C3: language processing, C4: computational psychiatry, C5: biology and morphogenesis). Classification uses word-
boundary-aware keyword matching against curated lists (74+ mathematical indicators, 25+ philosophy terms, 24+
tools terms, and 14–20 terms per application domain—totaling over 200 keywords across 8 categories, all documented
in config.yaml) applied to titles and abstracts. A priority system ensures that specific application domains (C1–C5,
priority 1) take precedence over tools (B, priority 2), formal theory (A1, priority 3), and the broad qualitative philoso-
phy catch-all (A2, priority 4). Within a priority tier, the domain with the most keyword matches wins. A1’s keyword
set includes mathematical indicators such as theorem, proof, convergence, posterior, equation, and Fokker–Planck, en-
suring that papers with mathematical content are classified as formal theory rather than defaulting to the philosophy
category.
3.5.2
Temporal Metrics and Growth-Rate Estimation
We compute temporal publication metrics including year-by-year counts with gap-filling, cumulative totals, 3-year
smoothed moving averages, and peak year identification. Field dynamics are estimated via two complementary metrics.
The mean year-over-year growth rate
̄𝑔is the arithmetic mean of annual growth rates for years with non-zero
prior-year publications. The doubling time 𝑡𝑑= ln 2/ ln(1 +
̄𝑔). The compound annual growth rate (CAGR)
captures the annualized rate across the full temporal span. Mathematical details are provided in Appendix 8.3.
3.5.3
Text Analytics
We construct the TF-IDF matrix using tokenization with stopword removal and L2-normalized smoothed term-
frequency inverse-document-frequency weighting [Salton et al., 1975], with a configurable vocabulary size (default:
500 features). We apply non-negative matrix factorization (NMF) to discover latent topics using multiplicative up-
date rules [Lee and Seung, 1999]. Topic count 𝑘= 5 was selected via expert-driven assessment. Mathematical details
are provided in Appendix 8.2.
3.5.4
Citation Network Construction
We construct the intra-corpus citation network as a directed graph where nodes are papers and edges represent citation
relationships resolved within the corpus. Because identifier formats vary across databases (arXiv IDs, DOIs, Semantic
Scholar IDs), only references whose identifiers match a corpus entry contribute edges; the resulting resolution rate
(7.4%) represents a lower bound on the true intra-corpus citation density. Network metrics include PageRank centrality,
HITS hub and authority scores [Kleinberg, 1999], degree distributions, network density, connected components, and
community structure via greedy modularity maximization [Clauset et al., 2004].
14

## Page 15

3.6
Stage 3: Nanopublication-Based Knowledge Graph
Stage 3 is the methodological core of this work: it transforms unstructured abstracts into a structured, RDF-compatible
knowledge graph of scientific evidence. The stage encompasses four tightly coupled operations: LLM-based assertion
extraction, nanopublication packaging, knowledge graph construction, and citation-weighted hypothesis scoring.
3.6.1
LLM-Based Assertion Extraction
We extract assertions by prompting a locally hosted LLM (Ollama [Ollama Team, 2024]) to assess each paper’s abstract
against eight standard hypotheses. The model receives a structured prompt containing the paper title, abstract, and
hypothesis definitions, and returns a JSON array where each element specifies a hypothesis ID, direction (supports,
contradicts, neutral, or irrelevant), a confidence score 𝑐∈[0, 1], and a reasoning string. Assertions marked “irrelevant”
are discarded; confidence values are clamped to [0, 1]; and responses are validated against the known hypothesis ID
set. Papers lacking abstracts are skipped. Detailed prompt engineering, error taxonomy, and validation methodology
are documented in the extraction pipeline section.
3.6.2
Nanopublication Schema and RDF Structure
Each assertion is encoded as a nanopublication [Groth et al., 2010, Kuhn et al., 2016]—a minimal, self-contained,
machine-readable unit of scientific evidence.
Formally, each nanopublication is a tuple (𝑝, ℎ, 𝑑, 𝑐) where 𝑝is the
paper identifier, ℎthe hypothesis identifier, 𝑑∈{supports, contradicts, neutral} the direction, and 𝑐the confidence.
Provenance metadata records the LLM model, UTC timestamp, and paper identifier.
The pipeline serializes nanopublications in two complementary formats:
1. JSON Lines (one JSON object per line) for efficient incremental checkpointing.
Assertions are saved at
configurable intervals (default:
every 50 papers), enabling the pipeline to resume from where it left off
after interruption without re-processing already-analyzed papers.
Deduplication uses the composite key
(𝑝𝑎𝑝𝑒𝑟_𝑖𝑑, ℎ𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠_𝑖𝑑); re-runs with improved models overwrite stale results.
2. RDF/TriG per the nanopublication standard (nanopub.net), producing four named graphs per nanopublication:
Table 3: RDF/TriG nanopublication structure.
Each nanopublication contains four named graphs encoding the
assertion, its provenance, and publication metadata per the nanopublication standard (nanopub.net).
Named Graph
Content
Key Predicates
Head
Links the nanopub resource to its three component graphs
np:hasAssertion, np:hasProvenance, np:has
Assertion
The core scientific claim
aif:asserts, aif:supports/aif:contradict
Provenance
How the assertion was generated
prov:wasGeneratedBy, prov:generatedAtTim
Publication Info
Metadata about the nanopublication itself
dc:created, dc:creator, dc:license
The namespace http://activeinference.institute/ontology/ (prefix aif:) defines all domain predicates; the
nanopublication schema (http://www.nanopub.org/nschema#, prefix np:)
provides structural predicates; prove-
nance uses PROV-O (http://www.w3.org/ns/prov#); and Dublin Core (http://purl.org/dc/terms/) provides
publication metadata. The TriG output is suitable for publication to the decentralized nanopublication network and
aligns with FAIR data principles: Findable via URI-based identification, Accessible via standard RDF protocols,
Interoperable through W3C-standard serialization, and Reusable with explicit provenance and CC0 licensing.
3.6.3
Knowledge Graph Construction
The knowledge graph is an RDF-compatible directed graph with three node types: paper nodes (metadata: title,
abstract, authors, year, venue, citation count, domain), assertion nodes (claim text, direction, hypothesis ID, confi-
dence), and hypothesis nodes (the eight standard hypotheses). Edges encode five relations defined in the schema:
• aif:asserts — Paper →Assertion
• aif:cites — Paper →Paper
• aif:belongsTo — Paper →Subfield
• aif:supports — Assertion →Hypothesis
• aif:contradicts — Assertion →Hypothesis
15

## Page 16

The graph is implemented with a dual backend: rdflib [RDFLib Team, 2023] when available (preferred for semantic
web compatibility), with automatic fallback to networkx.DiGraph for environments without RDF dependencies. Both
backends maintain identical internal indices for efficient paper, assertion, and hypothesis queries.
3.6.4
Citation-Weighted Hypothesis Scoring
For each hypothesis 𝐻, we compute a citation-weighted evidence score:
score(𝐻) =
∑𝑎∈𝑆(𝐻) 𝑤(𝑎) −∑𝑎∈𝐶(𝐻) 𝑤(𝑎)
∑𝑎∈𝐴(𝐻) 𝑤(𝑎)
(1)
where 𝑆(𝐻), 𝐶(𝐻), and 𝐴(𝐻) are the sets of supporting, contradicting, and all assertions for 𝐻, and the weight
function is:
𝑤(𝑎) = log(1 + citations(𝑎)) ⋅confidence(𝑎)
(2)
The logarithmic citation weighting ensures that highly cited papers carry more influence without allowing any single
paper to dominate. The score lies in [−1, 1]. Interpretation note: a score of +0.7 indicates that 70% of weighted
evidence supports the hypothesis (net of contradictions and normalized by total weighted evidence), not that the
hypothesis has a 70% probability of being true. Scores are best interpreted as relative rankings across hypotheses
and as temporal trajectories within a hypothesis, rather than as absolute probability estimates. Temporal trends are
computed by evaluating the cumulative score at each year, using only assertions from papers published up to that
year. A full derivation appears in Appendix 8.1.
3.6.5
Tally-Based Evidence Aggregation
We emphasize that this algorithmic scoring formula constitutes a tally-based approach to evidence synthesis: each
nanopublication assertion operates as an independent evidential vote, weighted by citation impact and the extraction
model’s confidence. The aggregation is linear and additive—supporting and contradicting assertions are summed and
differenced without modeling dependencies, correlated evidence, or causal structure among claims. This design choice
prioritizes transparency, reproducibility, and computational tractability over statistical sophistication.
The tally-based framing introduces three constraints. First, assertions from methodologically related papers (e.g.,
iterative publications from a single research group testing the same model) are counted independently, amplifying
correlated evidence. To illustrate: if a group publishes three papers (2019, 2021, 2023) reporting successively refined
variants of the same predictive coding model, each with high citation counts, the scoring formula counts three in-
dependent supporting assertions for H4—even though the underlying empirical evidence is largely overlapping. An
evidential diversity index (proposed in the conclusion) would downweight this cluster. Second, the scoring metric
treats all assertion sources symmetrically: an assertion from a theoretical review and one from an empirical trial
carry equal weight at a given confidence level. Third, temporal scoring tracks cumulative totals rather than dynamic
probabilistic estimates; the score at year 𝑡is the sum of all historical evidence, rather than a decaying posterior that
downweights early work.
We embrace these constraints intentionally. The tally-based approach provides a stable, interpretable baseline against
which more sophisticated scoring methods can be evaluated. The conclusion describes concrete extensions—including
hierarchical Bayesian scoring, causal evidence graphs, and evidential diversity indices that downweight correlated
evidence.
16

## Page 17

3.7
Stages 4–5: Visualization, Variable Injection, and Reproducibility
3.7.1
Stage 4: Visualization
Stage 4 renders 16 publication-ready figures from the analysis outputs of Stages 2 and 3. All figures use the Wong (2011)
colorblind-safe palette [Wong, 2011] and enforce a 16-point minimum font size for accessibility compliance. Figures
span six categories: field summary and domain distribution (2 figures), growth and temporal dynamics (2 figures),
citation network topology (2 figures), hypothesis evidence dashboard and timeline (2 figures), assertion composition
(2 figures), and text analytics—word cloud, PCA embeddings, term heatmap, dendrogram, topic-term bars, and
co-occurrence matrix (6 figures). The figure generation script reads only JSON outputs and produces only PNG
files, enforcing a strict, unidirectional data flow that prevents visualization operations from inadvertently modifying
analytical results.
3.7.2
Stage 5: Manuscript Variable Injection
Stage 5 computes dynamic variables from all pipeline outputs and injects them into manuscript Markdown templates
via double-brace placeholder substitution of the form {<>} wrapping a variable name (e.g. the literal token spelled
{{<CORPUS_SIZE>}} becomes the rendered corpus count). Variables include corpus-level metrics (size, year range,
CAGR), per-domain counts and percentages, citation network statistics (nodes, edges, density, components, resolution
rate, mean in-degree), hypothesis scores, and figure counts. All formatting (comma thousand separators, escaping)
is applied during variable computation, ensuring the manuscript templates remain human-readable while producing
publication-ready output.
Unrecognized placeholders are preserved with a warning logged, enabling incremental
manuscript development ahead of full pipeline execution.
3.7.3
Reproducibility and Test-Driven Validation
The pipeline is deterministic given fixed random seeds and API responses. Test-driven development enforces 90%
minimum code coverage on project modules and 60% on shared infrastructure, with real data and computation (no
mocking). The test suite validates boundary conditions for hypothesis scoring (all-support →+1, all-contradict →
−1, balanced →0), schema consistency, serialization round-trips, and end-to-end pipeline integrity.
Source code,
configuration, and outputs are available under CC-BY-4.0.
17

## Page 18

4
Results
4.1
Hypothesis Evidence Landscape and Temporal Dynamics
The LLM-based extraction pipeline produced a total of 1,490 assertions across the eight tracked hypotheses, drawn from
the full corpus of 𝑁= 819 papers. Before presenting the results, we reiterate the interpretive framework established
in the methodology: hypothesis scores are relative rankings among hypotheses and temporal trajectories within each
hypothesis—they are not absolute probability estimates. Publication bias and linguistic asymmetry (§4.1.4.1) inflate
all scores toward the positive end, and the tally-based aggregation does not model evidential dependencies.
The
distribution of assertion types and the resulting citation-weighted scores reveal a differentiated evidence landscape
(Figure 1):
Table 4: Citation-weighted hypothesis evidence landscape (𝑁= 819 papers, 1,490 total assertions).
Scores are
computed via (1) and range from −1 (unanimous contradiction) to +1 (unanimous support). “Character” summarizes
the qualitative evidence profile for each hypothesis.
Hypothesis
Score
Supports
Neutral
Contradicts
Total
Character
H7: Morphogenesis
+1.00
24
0
0
24
Strong consensus
H2: AIF Optimality
+0.97
291
6
2
299
Strong consensus
H4: Predictive Coding
+0.94
233
24
1
258
Strong consensus
H6: Clinical Utility
+0.93
23
2
0
25
Strong consensus
H5: Scalability
+0.85
108
13
0
121
Strong consensus
H8: Language AIF
+0.83
31
3
0
34
Strong support
H3: Markov Blanket Realism
+0.78
12
2
4
18
Moderate, active debate
H1: FEP Universality
+0.48
281
429
1
711
Broad but diffuse
Figure 1: Hypothesis scoring dashboard showing citation-weighted evidence scores ([−1, +1]) for the eight tracked
hypotheses, sorted descending by consensus strength. Predominantly positive scores reflect both genuine empirical
support and systematic positive biases from publication selection and linguistic framing (see §4.1.4.1).
4.1.1
Interpretation of Evidence Profiles
To directly address our core research questions—identifying which claims are robustly supported and which remain
contested—we evaluated how the hypothesis-level evidence maps against the critiques introduced in §3. The eight
18

## Page 19

hypotheses cluster into three tiers, defined by score ranges that emerge from the data rather than being imposed a priori.
The consensus tier (score > 0.83; H7, H2, H4, H6, H5) spans five of the eight hypotheses, revealing a predominantly
supportive evidence landscape across domains. H8 (Language AIF) sits at the boundary of consensus at score
+0.83 — above 0.8 but below the more stringent 0.83 line that separates the densely populated upper tier from
the rest. Morphogenesis (H7) achieves the maximum score (+1.00), though its small evidence base (24 assertions)
means unanimity reflects limited assessment scope rather than mature empirical closure. AIF Optimality (H2) holds
the second highest score (+0.97) despite carrying the largest raw count of contradicting assertions (2): supporting
assertions are substantially more highly cited than critical ones, so citation-weighting amplifies the supportive signal—
underscoring that citation-weighted scores capture which claims the community cites most, not a simple ballot of
assertion counts. Predictive coding (H4), the most extensively assessed hypothesis with 258 assertions and a score of
+0.94, has accumulated overwhelmingly supportive evidence since the 1970s, reflecting the deep empirical grounding
of hierarchical prediction error models in neuroscience. This trajectory is consistent with the manual benchmarking
results of Knight et al. [Knight et al., 2022], which similarly identified predictive coding as the most rigorously validated
construct in the corpus. Clinical Utility (H6, +0.93) and Scalability (H5, +0.85) complete the upper consensus tier;
H5’s trajectory accelerated sharply after 2017 as deep active inference architectures emerged. The H8 (Language AIF)
boundary placement noted above reflects recent breakthroughs coupling active inference to large language models
within a still-maturing evidence base.
The moderate tier (score 0.5–0.8; H3) contains a single hypothesis. Markov blanket realism (H3) has the small-
est overall evidence base (18 assertions) with a score of +0.78 and 4 contradicting assertions—empirically captur-
ing the ongoing philosophical debate between those who treat Markov blankets as real thermodynamic boundaries
(Friston blankets'') and those who argue they are purely instrumental statistical tools (Pearl blan-
kets’ ’) [Bruineberg et al., 2022]. The moderate score for H3 reflects this active ontological debate: the supporting
literature is more highly cited but not by a large margin, and the small total evidence base limits inferential confidence.
The diffuse tier (score < 0.5; H1) is the most diagnostically informative for understanding the field’s intellectual
maturation. FEP universality (H1) generates one of the largest raw evidence bases (711 assertions) yet achieves a score
of only +0.48—a striking gap explained by assertion composition: neutral assessments account for 429 of those 711
tallies, while supporting assertions number 281 and contradicting assertions just 1. This neutral plurality—more than
either supporting or contradicting tallies—reveals that researchers routinely invoke the FEP as conceptual scaffolding
without subjecting its universality claim to explicit empirical test. This composition is the quantitative fingerprint
of the falsifiability critique leveled by Colombo and Seri‘es [Colombo and Seriès, 2021]: a principle elastic enough
to accommodate any self-organizing system without generating predictions that distinguish it from alternatives will
naturally accumulate invocations rather than tests, and invocations register as neutral in the extraction pipeline.
4.1.2
Temporal Dynamics of Evidence Accumulation
The cumulative evidence timeline (Figure 2) reveals three temporal patterns. First, early convergence: H4 (predic-
tive coding) reached positive territory in the late 1990s following the publication of Rao and Ballard’s foundational
predictive coding model [Rao and Ballard, 1999] and has maintained a high score since, reflecting the mature em-
pirical base in cognitive neuroscience. Second, recent acceleration: H5 (scalability) and H6 (clinical utility) show
steep upward trends after 2017, tracking the emergence of deep active inference tools and computational psychiatry
applications. The H5 trajectory reflects a cumulative body of work culminating in benchmark demonstrations such
as AXIOM [Heins et al., 2025], which showed that object-centric world models under AIF can match state-of-the-art
deep RL performance—but the temporal trend was already positive before any single result, and the score captures
the aggregate rather than any individual paper. Third, moderate and stable: H3 (Markov blanket realism) has
maintained a score in the moderate range since 2018, with supporting papers partially offset by targeted philosophical
critiques—a pattern consistent with ongoing debate rather than either clear consensus or rejection.
4.1.3
Assertion Composition and Distribution
The per-hypothesis composition of assertions (Figure 3) and the multi-panel summary (Figure 4) provide complemen-
tary views of the extraction results.
4.1.4
Limitations of the Current Scoring Approach
4.1.4.1
Publication Bias and Linguistic Asymmetry
The predominantly positive scores observed across all
eight hypotheses should be interpreted with two systematic caveats.
19

## Page 20

Figure 2: Temporal evolution of cumulative citation-weighted evidence scores by hypothesis (2005–2026). Divergent
trajectories around the shaded neutral boundary (±0.1) reveal which hypotheses are gaining or losing support over
time. H4 (predictive coding) stabilized early; H5 (scalability) accelerated post-2017.
Figure 3: Stacked horizontal bars decomposing per-hypothesis assertions into supports (green), contradicts (red-
orange), and neutral (blue) categories (𝑁= 1, 490 total assertions). Labels show total count and support percentage.
The high support fractions are partially attributable to publication bias and affirmative linguistic framing.
20

## Page 21

Figure 4:
Multi-panel assertion summary:
(left) pie chart of overall assertion type distribution showing sup-
ports/contradicts/neutral proportions, (right) per-hypothesis assertion counts with palette-coded bars. 𝑁= 1, 490
assertions extracted from 819 papers.
First, publication bias systematically inflates supporting evidence. Academic journals preferentially publish positive
and confirmatory results (Sterling 1959), meaning that studies finding null or contradictory outcomes for any hypothesis
are less likely to appear in the retrievable literature.
This file-drawer effect is well-documented across scientific
disciplines and is expected to disproportionately suppress contradicting assertions in our extraction pipeline. The
Active Inference literature is particularly susceptible: as a theoretical framework with strong foundational proponents,
papers are more likely to frame results as consistent with the FEP than as challenges to it.
Second, linguistic asymmetry in academic writing further skews extraction toward positive classifications. Declar-
ative scholarly claims are inherently phrased affirmatively—authors write “our results support,” “consistent with,” or
“extends the prediction of” far more frequently than “our results refute” or “contradicts the claim that.” Because the
LLM extraction pipeline operates on abstract text, this linguistic imbalance propagates directly into the assertion
distribution. Even papers presenting genuinely mixed evidence tend to frame their abstracts in terms of what was
found rather than what was not, biasing the extracted direction toward “supports.’ ’
These two effects act in concert: publication bias reduces the number of contradicting papers in the corpus, and
linguistic framing reduces the number of contradicting assertions extracted from the papers that do appear. Conse-
quently, the absolute values of hypothesis scores should not be taken as unbiased measures of scientific consensus.
The relative ordering and temporal trajectories of hypothesis scores are more robust indicators, as these biases affect
all hypotheses approximately equally.
4.1.5
Methodological Validation and LLM Calibration
The evidence derives from automated LLM-based assertion extraction operating on abstracts only, without human-
validated ground truth calibration; confidence scores are self-assessed and uncalibrated; the pipeline uses 𝑐≥0.60
threshold to mitigate over-extraction. Relative rankings are more robust than absolute scores. A formal validation
protocol (10% manual annotation, Cohen’s 𝜅, boundary-case auditing) remains a critical next step.
21

## Page 22

4.2
Field Overview: Disciplinary Structure and Growth Dynamics
Annual output in the Active Inference literature rose from 1 papers in 2005, reaching a peak of 123 papers in 2025—
a transition from a niche within theoretical neuroscience to a multi-disciplinary research program spanning three
primary domains and eight tracked categories. The corpus start of 2005 was chosen to capture Energy-Based Model
and variational Bayesian antecedents [Dayan et al., 1995, LeCun et al., 2006] that preceded the formal introduction of
the Free Energy Principle in 2006 [Friston et al., 2006] and its subsequent full elaboration [Friston, 2010]. Our corpus,
extracted from arXiv, Semantic Scholar, and OpenAlex and deduplicated to 𝑁= 819 papers (2005–2026), captures
the breadth, tempo, and internal architecture of this expansion (Figure 5).
Figure 5: Publication counts by domain (𝑁= 819). Application domains (C1–C5) collectively account for the largest
share of the corpus; Domain A2 (qualitative philosophy) is the largest single category, reflecting the FEP’s broad
theoretical reach.
4.2.1
Corpus-Level Summary
Table 5: Corpus-level summary statistics for the Active Inference literature corpus (𝑁= 819), spanning 2005–2026.
Metric
Value
Total papers
819
Year range
2005–2026
Peak year
2025
CAGR
20.36%
Active domains
8 of 8 tracked (A1–A2, B, C1–C5)
The CAGR of 20.36% (measured as the annualised growth rate of yearly publication volume between endpoint years
2005 and 2026) reflects sustained field expansion; the actual rapid growth phase began around 2013, with annual
output accelerating substantially (Figure 6). Sustained high output persisting into subsequent years suggests the
field has reached a mature production phase rather than experiencing a transient spike. Citation network metrics are
detailed in the dedicated citation network analysis (see the citation network analysis).
4.2.2
Domain Distribution
Keyword-based classification assigns each paper to one of eight categories across three domains:
22

## Page 23

Figure 6: Annual (bars) and cumulative (line) publication counts, 2005–2026 (𝑁= 819, CAGR = 20.36%). The
inflection around 2013 marks the onset of rapid growth. Moving average trendline (dashed), peak year, and median
year annotated.
Table 6: Domain distribution across three tiers and eight categories (𝑁= 819 papers). Classification uses hierarchical
keyword matching with priority-based routing to minimize over-assignment to catch-all categories.
Domain
Category
Papers
Percentage
A – Core Theory
A1: Formal Theory
64
7.8%
A2: Qualitative Philosophy
60
7.3%
B – Tools
B: Tools & Translation
170
20.8%
C – Applications
C1: Neuroscience
161
19.7%
C2: Robotics
136
16.6%
C3: Language
58
7.1%
C4: Psychiatry
33
4.0%
C5: Biology
135
16.5%
The concentration of papers in A2 (qualitative philosophy and general theory) reflects the broad scope of foundational
FEP work (Figure 7). The priority-based classifier mitigates over-assignment by routing papers with mathematical
indicators (theorems, proofs, equations, statistical formalism) to A1 before falling back to A2, and by preferring
specific application domains (C1–C5) and tools (B) over both core-theory categories. Papers that discuss FEP/AIF
conceptually without mathematical formalism or domain-specific vocabulary are correctly assigned to A2. This figure
should be read as a ceiling on theoretical generality rather than a literal measure of research focus—embedding-
based classification would likely redistribute some fraction into more specific categories. That all eight categories
are populated, including computational psychiatry (C4) and formal theory (A1), indicates diversification beyond the
field’s neuroscience origins.
Detailed characterizations of each domain—including historical context, growth trends, and open problems—are pro-
vided in the supplementary domain analyses (see the domain analyses). Latent topic structure, vocabulary analysis,
and document embeddings are presented in the text analytics section (see the text analytics section).
4.2.3
Cross-Domain Comparison
Three structural features emerge from the cross-domain comparison (Figure 8). First, no single legacy domain dom-
inates: Domain B (Tools & Translation) accounts for 20.8% of the corpus, followed by C1 (Neuroscience) at 19.7%
and C2 (Robotics) at 16.6%. Second, Domain A (Core Theory) aggregates 15.2% collectively (A1 + A2), while the
23

## Page 24

Figure 7: Domain distribution (𝑁= 819). Classification uses hierarchical keyword matching against curated lists
applied to titles and abstracts, capturing distinct methodological and domain-specific groupings.
Table 7: Cross-domain comparison showing growth trajectories, maturity levels, key challenges, and representative
publications for each of the eight tracked categories. Growth trends and maturity assessments are based on temporal
publication patterns and evidence base depth.
Domain
Category
Papers
Growth
Maturity
Key Challenge
Rep. Work
A
A1: Formal
64 (7.8%)
Growing
Mature
Math accessibility
[Sakthivadivel, 2023]
A
A2: Philosophy
60 (7.3%)
Stable
Mature
Catch-all absorption
[Friston, 2010]
B
B: Tools
170 (20.8%)
Rapid
Growing
Deep RL benchmarks
[Fountas et al., 2020]
C
C1: Neuroscience
161 (19.7%)
Stable
Mature
Theory–neuroimaging gap
[Clark, 2013]
C
C2: Robotics
136 (16.6%)
Growing
Growing
Embedded real-time
[Lanillos et al., 2021]
C
C3: Language
58 (7.1%)
Emerging
Nascent
NLP model comparison
[Friston et al., 2020]
C
C4: Psychiatry
33 (4.0%)
Emerging
Nascent
Clinical translation
[Smith et al., 2022]
C
C5: Biology
135 (16.5%)
Rapid
Nascent
Empirical validation
[Kuchling et al., 2020]
24

## Page 25

emergent application frontiers (C3–C5) exhibit accelerating growth. Third, A1’s 64 papers understate its intellectual
influence—the mathematical formalisms developed in A1 shape implementations across all domains.
Figure 8: Stacked area chart of publications by domain, 2005–2026 (𝑁= 819). A2 (qualitative philosophy) provides
a large baseline; application domains C1–C5 show accelerating diversification from 2015 onward.
25

## Page 26

4.3
Domain Analyses: Growth Trajectories and Open Problems
This supplementary section provides detailed characterizations of each of the eight tracked Active Inference domains,
organized under three tiers: A (Core Theory), B (Tools & Translation), and C (Application Domains).
4.3.1
Domain A: Core Theory
4.3.1.1
A1 — Quantitative & Formal Theory (𝑛= 64, 7.8%)
The A1 domain develops the mathematical
foundations underpinning the Free Energy Principle: information geometry, category-theoretic formulations of Markov
blankets, path integral formulations of free energy minimization, and gauge-theoretic perspectives on self-organization.
A central debate concerns the ontological status of Markov blankets—whether they correspond to real physical bound-
aries or are merely useful statistical constructs [Bruineberg et al., 2022]. Bruineberg et al. draw a critical distinction
between Pearl blankets (instrumental, epistemic tools for conditional independence in Bayesian networks) and Friston
blankets (ontologically laden physical boundaries between agent and environment), arguing that the scientific credi-
bility of the former should not be extended uncritically to the latter. Friston and collaborators continue to address
this critique through the development of Bayesian mechanics [Sakthivadivel, 2023], which aims to place the FEP
on firmer mathematical footing by grounding Markov blanket dynamics in the physics of belief-based systems. Our
hypothesis scoring quantifies this debate: the Markov blanket realism hypothesis (H3) achieves a score of +0.78 with
4 contradicting assertions, making it the most heavily contested hypothesis in the corpus. Recent theoretical consoli-
dation has strengthened the formal tools available to A1: variational message passing formulations [Champion et al.,
2021] connect expected free energy decomposition—into risk, ambiguity, epistemic, and instrumental components—to
practical planning algorithms, advancing the theoretical justification for EFE-based policy selection. Path integral
formulations now connect Markov blanket dynamics to least-action principles, framing free energy minimization as
paths of least action for belief updating. With 64 papers (7.8% of the corpus), A1 captures a meaningful share of
formal work, reflecting the improved classifier’s ability to route papers with mathematical formalism (theorems, proofs,
convergence, posterior distributions, Fokker–Planck equations) into this domain rather than the qualitative philosophy
catch-all. Key evidence gap: A mathematically formal distinction yielding testable predictions that differentiate
systems actively minimizing an internal free energy functional from systems that merely possess a Markov blanket.
4.3.1.2
A2 — Qualitative Philosophy & General Theory (𝑛= 60, 7.3%)
The A2 domain encompasses
papers that develop, extend, or review the core Free Energy Principle and Active Inference framework without re-
stricting attention to a specific application domain. This includes Friston’s foundational work on variational free
energy minimization [Friston, 2010], the textbook treatment by Parr, Pezzulo, and Friston [Parr et al., 2022], and
numerous tutorial and review papers. The priority-based classifier mitigates over-assignment to A2 by routing papers
with mathematical formalism to A1 and papers with domain-specific vocabulary to C1–C5 or B before the A2 catch-all
is reached. Nevertheless, the count likely still conceals meaningful internal structure: papers addressing embodied
cognition, Bayesian brain theory, and philosophical implications of the FEP are all subsumed under this heading.
Three unresolved debates drive the most contested A2 literature. First, the explanatory scope question: is the
FEP a principle of physics (applying to any system at non-equilibrium steady state [Friston, 2010]), a principle of
biology (restricted to organisms that actively maintain their boundaries against entropy), or a computational-level
description of cognition [Clark, 2013]? The answer determines whether evidence from robotics, synthetic biology, or
cellular dynamics counts as genuine support for the FEP or merely analogical illustration. Second, the relationship
to reinforcement learning: active inference and deep RL both minimize expected future cost, but differ in whether
the objective is expected free energy (AIF) or expected cumulative reward (RL). Establishing formal equivalence or
principled divergence between these frameworks is prerequisite for the benchmark comparisons domain B requires.
Third, eliminativist vs. instrumentalist interpretations of free energy itself—whether variational free energy is
a latent quantity the brain actually tracks or a mathematical convenience for describing inference—remain open, with
consequences for the empirical status of A1 formalisms. Key evidence gap: A head-to-head theoretical comparison
showing conditions under which active inference makes predictions that differ from reinforcement learning, optimal
control, or Bayesian brain models, together with experimental designs capable of adjudicating among them.
4.3.2
Domain B: Tools & Translation Methods
4.3.2.1
B — Algorithms, Scaling, and Software (𝑛= 170, 20.8%)
Domain B addresses the computational
challenge of making active inference practical in complex, high-dimensional environments. Early implementations
relied on small discrete state spaces amenable to exact message passing. Recent work has introduced deep active
inference using neural networks to amortize inference [Fountas et al., 2020], Monte Carlo tree search for planning
[Champion et al., 2021], hybrid architectures combining model-based planning with model-free components, and
26

## Page 27

interpretable alternatives such as Free Energy Projective Simulation (FEPS) [Pazem et al., 2024], which exposes
decision logic as human-readable policy graphs. The central open question is whether active inference agents can
match deep reinforcement learning performance on standard benchmarks while retaining interpretability and sample
efficiency. The availability of the pymdp library [Heins et al., 2022] has lowered implementation barriers, contributing to
this domain’s growth. The recent establishment of the Pymdp Fellowship program (funding 8 open-source developers
in 2025) and the release of real-time stream processing tools like RxInfer.jl v4.0.0 [Bagaev et al., 2025] indicate a
vibrant and maturing software ecosystem. Key evidence gap: Head-to-head benchmarking of AIF agents against
state-of-the-art deep RL baselines on standardized, continuous-control or long-horizon environments.
4.3.3
Domain C: Application Domains
4.3.3.1
C1 — Neuroscience (𝑛= 161, 19.7%)
Neuroscience represents the historical core of the Active Infer-
ence research program. The predictive processing account—in which cortical hierarchies minimize prediction errors
through both perceptual inference and active sampling—remains one of the most empirically tested aspects of the
framework [Friston, 2010, Clark, 2013]. The broader neuroscience literature on Dynamic Causal Modeling and predic-
tive coding is extensive; the relatively modest count here likely reflects the keyword classifier’s inability to distinguish
neuroscience-specific applications from general FEP theory. Bridging the gap between computational models and
empirical neuroimaging data remains the domain’s primary challenge.
4.3.3.2
C2 — Robotics (𝑛= 136, 16.6%)
Robotics applications treat embodied agents as free energy minimizing
systems that unify perception and action through proprioceptive and exteroceptive prediction errors [Lanillos et al.,
2021].
Applications include robotic arm control, mobile navigation, manipulation, and multi-robot coordination.
Active inference offers roboticists a principled framework for integrating sensory processing, motor planning, and
adaptive behavior without separate perception and control modules. Key challenges include real-time computational
feasibility on embedded hardware, continuous high-dimensional action spaces, and sim-to-real transfer.
4.3.3.3
C3 — Language Processing (𝑛= 58, 7.1%)
The C3 domain conceptualizes linguistic processes—
speech perception, sentence comprehension, dialogue, and reading—as active inference operating over deep hierarchical
generative models of linguistic structure [Friston et al., 2020]. Active inference models of reading have reproduced
saccadic eye-movement patterns, while models of speech perception capture how listeners integrate prior expectations
with acoustic evidence. Recent work couples active inference to large language models, pragmatics, and multi-agent
communication. The connection between AIF and LLMs runs in both directions: Wen [Wen, 2025] proposes that AIF
can replace external reward signals in LLM-based agents, while Friston et al. [Friston et al., 2025] demonstrate how
active inference enables artificial reasoning through structure learning via Bayesian Model Reduction. The language
domain is also where AIF shows strong results through novel discrete generative models for structured sequential tasks
[Millidge, 2024].
4.3.3.4
C4 — Computational Psychiatry (𝑛= 33, 4.0%)
Computational psychiatry leverages active inference
to model psychiatric conditions as disruptions in belief updating, precision weighting, or prior rigidity [Smith et al.,
2022]. Schizophrenia has been modeled as impaired precision weighting on bottom-up prediction errors; depression
as over-precise negative priors; and autism spectrum conditions as atypical precision allocation over sensory channels.
Beyond clinical psychopathology, the framework is now being extended to model higher-order cognition: Whyte et
al. [Whyte et al., 2025] propose a metacognitive active inference account of imaginative experience, in which “inner
screen” representations emerge from EFE-driven attention allocation under FEP constraints—connecting computa-
tional psychiatry to consciousness research. The domain continues to expand, with emerging frameworks integrating
psychodynamic theory (e.g., self-identity formation via embodied interactions) with predictive processing to unify en-
vironmental and biological factors underlying stress disorders. Translating these computational models into diagnostic
markers and therapeutic protocols remains an ongoing challenge. Key evidence gap: Translating retrodictive com-
putational phenotyping models into prospective clinical predictions that demonstrably outperform standard diagnostic
criteria in clinical trials.
4.3.3.5
C5 — Biology & Morphogenesis (𝑛= 135, 16.5%)
The C5 domain applies active inference and the
FEP to biological systems beyond the brain: cellular behavior, morphogenesis, evolutionary dynamics, and the origins
of life. Morphogenetic processes have been modeled as collective active inference, where groups of cells coordinate to
minimize a shared free energy functional [Kuchling et al., 2020, Levin, 2022]. Recent empirical work has validated
collective AIF at larger scales: Heins et al. [Heins et al., 2024] demonstrated that surprise minimization alone produces
realistic collective motion patterns, providing a principled alternative to ad hoc flocking rules.
The FEP’s reach
27

## Page 28

now extends beyond biological organisms into engineered systems: Nazemi et al. [Nazemi et al., 2025] apply active
inference to smart building energy control under partial observability and privacy constraints, demonstrating that the
free energy framework can govern resource allocation in cyber-physical systems. As the second-largest domain, C5
reflects growing interest in extending the FEP to encompass all self-organizing systems—living and artificial—though
the ratio of theoretical proposals to empirical validation remains high.
4.3.4
Comparative Synthesis
Taken together, the three domains reveal a field transitioning from a focused neuroscience program to a broad inter-
disciplinary framework. The core–periphery structure is clear: Domain A provides the theoretical and mathematical
substrate, Domain B pursues engineering viability through scalable algorithms and software, and Domain C tests the
framework’s generality across neuroscience (C1), robotics (C2), language (C3), psychiatry (C4), and biology (C5). The
consistent pattern across applied domains—strong theoretical motivation paired with limited empirical validation—
suggests that the field’s next growth phase will depend on accumulating experimental evidence.
In direct response to RQ1 (How is the Active Inference field structured?), the domain taxonomy reveals an asymmetric
three-tier architecture: a dominant theoretical core (A), a growing translational layer (B), and an expanding but
empirically sparse application periphery (C). The keyword classifier’s heavy A2 concentration likely masks genuine
diversity within the theoretical core, but the architecture itself—theory →tools →applications—is robust across
classification approaches.
4.3.4.1
Domain–Hypothesis Cross-Reference
Each domain has a primary hypothesis linkage (see the detailed
hypothesis evidence analysis in the hypothesis results):
Table 8: Domain–hypothesis cross-reference linking each of the eight tracked categories to its primary hypothesis and
the direction of the current evidence base. See the hypothesis results for quantitative scores and temporal trends.
Table values are regenerated automatically from hypothesis_scores.json; the most recent verified pipeline run is
dated 2026-04-28.
Domain
Category
𝑛
Primary Hypothesis
Evidence Direction
A1
Formal
64
H3 Markov Blanket Realism
Contested
A2
Philosophy
60
H1 FEP Universality
Strongly supporting
B
Tools
170
H5 Scalability
Mixed
C1
Neuroscience
161
H4 Predictive Coding
Supporting
C2
Robotics
136
H2 AIF Optimality, H5 Scalability
Mixed
C3
Language
58
H8 Language AIF
Emerging
C4
Psychiatry
33
H6 Clinical Utility
Supporting
C5
Biology
135
H7 Morphogenesis
Supporting
The evidence directions summarized above are elaborated quantitatively—with citation-weighted scores, temporal
trends, and three-tier evidence profiling—in the hypothesis results section.
28

## Page 29

4.3.5
Text Analytics: Topic Modeling, Vocabulary Structure, and Document Embeddings
This section examines the latent semantic structure of the Active Inference corpus through complementary text-analytic
methods: non-negative matrix factorization for topic discovery, TF-IDF vocabulary analysis, document embedding
projections, and term co-occurrence patterns. Together, these analyses reveal thematic structure that cuts across the
keyword-based domain taxonomy presented in the field overview.
4.3.6
Topic Modeling: Latent Structure
Non-negative matrix factorization (NMF) applied to the TF-IDF matrix identifies five latent topics:
Table 9: Non-negative matrix factorization (NMF) topic decomposition of the corpus TF-IDF matrix (𝑘= 5 topics).
Top terms are ranked by NMF component weight; interpretations reflect dominant thematic content.
Topic
Top Terms
Interpretation
0
learning, agent, model, agents, active, environments, aif, inference, environment, based
Agent-environment
robotic applications
1
inference, active, energy, free, variational, control, bayesian, expected, optimal, principle
Active inference agen
making
2
states, internal, external, systems, markov, system, dynamics, information, beliefs, self
Markov blankets and
states
3
fep, systems, ai, principle, energy, free, theory, networks, modeling, language
Free energy principle
4
predictive, brain, cognitive, prediction, perception, processing, sensory, models, coding, model
Predictive coding and
science
4.3.6.1
Topic–Domain Overlap
These topics are partially orthogonal to the domain taxonomy. Topic 0 (agent-
environment modeling) spans tools (B), robotics (C2), and core theory (A1)—a cross-cutting theme that the keyword
classifier cannot capture. Topic 4 (predictive coding and cognitive neuroscience) aligns closely with neuroscience (C1)
but also draws from core theory. Topic 2 (Markov blankets and states) captures the mathematical core shared across
domains. Topic 3 (FEP and AI systems) reveals the growing intersection of active inference with mainstream artificial
intelligence research. The extracted topics demonstrate high stability; rerunning NMF across multiple random seed
initializations yields identical topic clusters (Jaccard similarity > 0.90 for top term sets). The absence of retrieval
noise (no spurious physics topics) confirms that the phrase-matched arXiv query effectively filters irrelevant content
(Figure 9).
Figure 9: Top 10 terms per NMF topic (𝑘= 5 topics, 500 vocabulary features). Term weights reflect NMF component
loadings; higher-weighted terms define each topic’s semantic focus.
29

## Page 30

4.3.7
Vocabulary Analysis
Figure 10: Word cloud of corpus vocabulary (𝑁= 819 abstracts) sized by maximum NMF component weight. Promi-
nent terms—”inference,” ”active,” ”free energy,” ”model”—reflect the field’s core theoretical commitments.
The word cloud (Figure 10) reveals the conceptual core of the Active Inference literature: terms related to the Free
Energy Principle (“inference,” “active,” “free energy,” “model,” “bayesian”) dominate, while application-specific terms
appear at smaller scales, reflecting the domain distribution’s heavy A2 concentration.
4.3.8
Document Embedding Projections
Principal Component Analysis of the TF-IDF document-term matrix projects each paper into a two-dimensional space
that preserves the directions of maximum variance (Figure 11). Rather than serving solely as a visual clustering aid,
this projection provides a quantitative measure of semantic distance between subfields.
The scatter plot, colored
by domain assignment, reveals the degree of semantic separation between domains. Loading arrows overlay the top-
variance terms, showing which vocabulary drives the principal components and highlighting the structural overlap
between theoretically similar domains that keyword-based hard categorization obscures.
4.3.9
Domain Semantic Similarity
To further interrogate the latent semantic structure of the subfields, we extract the top characterizing terms for
each domain and compute a hierarchical clustering of domain centroids. The heatmap (Figure 12) reveals distinctive
vocabulary patterns beyond mere keyword-level classification, while the dendrogram (Figure 13) confirms the tight
semantic proximity between Core Theory subfields (A1, A2) and the methodological alignment of Tooling (B) with
Robotics (C2).
4.3.10
Term Co-occurrence Patterns
The co-occurrence matrix (Figure 14) for the 30 most frequent corpus terms reveals tightly coupled term clusters
corresponding to the NMF topics. The strong co-occurrence between “free,” “energy,” “principle,” and “bayesian”
30

## Page 31

Figure 11: PCA projection of TF-IDF document embeddings (𝑁= 819 documents, 500 features), colored by domain.
Loading arrows indicate vocabulary terms contributing most to each principal component.
Variance explained is
annotated per axis.
anchors the theoretical core, while application-specific term clusters (e.g., “brain”–“cognitive”–“predictive”–“coding”)
form distinct off-diagonal blocks. The relative isolation of robotics-specific terms from neuroscience terms confirms
the semantic separation between these application domains despite their shared theoretical foundation.
31

## Page 32

Figure 12: Mean TF-IDF weight for the top 20 terms across all 8 domains. Darker cells indicate higher usage within a
domain, revealing distinctive vocabulary patterns beyond the keyword-level classification used for subfield assignment.
Figure 13: Hierarchical clustering of domain centroids (Ward linkage on mean TF-IDF vectors, 8 domains). Cophenetic
correlation annotated on figure. A1 (formal theory) and A2 (philosophy) cluster closely, as do C2 (robotics) and B
(tools).
32

## Page 33

Figure 14: Normalized co-occurrence matrix for the 30 most frequent terms across 𝑁= 819 abstracts. Cell intensity
reflects the fraction of documents in which two terms co-appear, normalized to [0, 1].
4.4
Citation Network Topology
The intra-corpus citation network provides a structural view of how Active Inference research is organized, identifying
influential hub papers, community structure, and patterns of citation isolation (Figure 15).
4.4.1
Network Density and Degree Distribution
The intra-corpus citation network contains 817 nodes and 2,176 edges, with a density of 0.33% and 547 connected
components. The average in-degree of ≈2.7 indicates that most papers receive few intra-corpus citations, consistent
with the field’s rapid expansion: the majority of recent papers have not yet accumulated citations within the corpus
(Figure 16). Only 7.4% of all identified references (2,176 intra-corpus matches out of 29,323 total reference entries)
resolve to other papers within the corpus, reflecting cross-source identifier mismatches and the field’s engagement with
a broad external literature base. Community detection identifies clusters via greedy modularity maximization [Clauset
et al., 2004].
4.4.2
Connected Components and Citation Isolation
The high number of connected components (547 out of 817 nodes) reveals that much of the corpus consists of citation-
isolated papers—works that neither cite nor are cited by other papers in the collection. A single Giant Connected
Component (GCC) typically dominates mature scientific networks; here, with 547 components across 817 nodes, the
GCC contains a minority of nodes while the remainder form singletons or small clusters of two to three papers. This
33

## Page 34

Figure 15: Intra-corpus citation network (𝑁= 819 nodes, 2,176 edges).
Node size reflects in-degree (number of
intra-corpus citations received); highly cited foundational papers serve as nexus points connecting sub-domains.
Figure 16: In-degree distribution of the citation network. The power-law tail is characteristic of citation networks,
with a small number of highly cited hubs.
34

## Page 35

is partially an artifact of cross-source identifier mismatches, but it also reflects the field’s pattern of papers engaging
with the FEP literature conceptually without building explicit, graph-tractable citation chains. PageRank analysis
identifies highly influential papers, predominantly Friston’s foundational work [Friston, 2010] and the AIF textbook
[Parr et al., 2022], which serve as nexus points linking otherwise disconnected subgraphs.
4.4.3
Network Summary
Table 10: Intra-corpus citation network summary statistics (𝑁= 819 papers). The low density and high component
count reflect the field’s rapid expansion and cross-source identifier mismatches.
Metric
Value
Nodes
817
Edges
2,176
Reference resolution rate
7.4% (2,176 / 29,323)
Connected components
547
Network density
0.33%
Mean in-degree
≈2.7
The citation topology corroborates the field overview findings (RQ1, RQ2): a small number of foundational papers—
predominantly Friston’s free energy and active inference formulations—anchor a rapidly expanding periphery of increas-
ingly specialized work. The extremely low density (0.33%) corresponds to an epistemic stage of high fragmentation,
meaning that literature synthesis and cross-pollination between specific sub-domains remain difficult. Theoretical
influence flows primarily through shared conceptual foundations (the hub nodes) rather than through dense mutual
citation across the periphery. As metadata standardization improves and DOI adoption becomes universal across
preprint and journal ecosystems, re-running this pipeline should yield substantially higher reference resolution rates
and a more connected graph, enabling finer-grained community detection and tracking.
35

## Page 36

5
Conclusion:
Evidence Landscape, Methodological Limitations, and
Research Agenda
5.1
Summary
This work demonstrates a first-generation prototype infrastructure for computational meta-analysis of a rapidly grow-
ing scientific field. By combining multi-source retrieval (𝑁= 819 papers from three databases), LLM-based assertion
extraction encoded as nanopublications, and citation-weighted hypothesis scoring, we produce a queryable, RDF-
compatible knowledge graph that tracks the evolving evidence for eight core Active Inference claims. The system
demonstrates the feasibility of automated living reviews, while clearly delineating the boundaries of current model
capabilities.
All assertions and hypothesis scores in this work are machine-generated without human validation. While the pipeline
is designed for rigor, these results should be treated as preliminary evidence requiring manual review before scientific
acceptance.
5.2
Constraints and Methodological Scope
Several conscious design constraints scope these findings.
5.2.1
Keyword Classifier Resolution
The keyword-based classifier operates over 200+ keyword indicators distributed across 8 domain categories (74 mathe-
matical indicators in A1 alone), using a deterministic priority system that routes papers to specific application domains
(C1–C5) before testing tools (B), formal theory (A1), and the qualitative philosophy catch-all (A2). Word-boundary-
aware matching reduces partial-match false positives, but keyword-based methods cannot capture semantic nuance:
papers using novel terminology or discussing cross-domain topics without standard vocabulary risk misclassification.
Residual A2 concentration should be interpreted as a ceiling on broad theoretical generality rather than a literal mea-
sure of philosophical focus. An embedding-based classifier trained on a labeled subset would provide a quantitative
upper bound on the fraction of A2 papers that merit redistribution.
5.2.2
Citation Network Coverage Gaps
The 2,176 intra-corpus edges spanning 547 connected components provide a topological skeleton, but three systematic
gaps inflate the component count: (1) cross-source identifier mismatches (DOI vs. OpenAlex vs. arXiv ID), (2) papers
whose references are not indexed by any source API, and (3) open-access preprints whose DOIs differ from their
published versions. Exhaustive DOI-level cross-matching with fuzzy title matching would condense the graph further.
5.2.3
Corpus Biases, Citation Dynamics, and Linguistic Framing
Citation counts are subject to Matthew effects and cumulative field-size biases. Partial-year indexing for the most
recent calendar year undercounts recent publications.
The measured 20.36% CAGR reflects the dilutive effect of
the long temporal span (2005–2026), corresponding to a 2.8-year publication doubling time; the growth phase from
2010 onward follows a steeper trajectory with a substantially shorter doubling interval. Additionally, the retrieved
corpus itself suffers from selection biases inherent to queried databases, including English-language dominance and
the structural over-indexing of preprints relative to peer-reviewed final versions. Finally, the predominantly positive
hypothesis scores across the board are inflated by two systematic effects: (1) publication bias, which causes academic
journals to preferentially select positive and confirmatory findings [Sterling, 1959], and (2) linguistic asymmetry
in scientific writing, where declarative claims are phrased affirmatively far more often than negatively. These effects
jointly suppress contradicting assertions in the extracted evidence base. Relative rankings and temporal trajectories
are more reliable than absolute score magnitudes.
5.2.4
LLM Extraction Fidelity, Domain Drift, and Robustness
Zero-shot LLM extraction introduces distinct systematic biases: over-extraction (the model hallucinating certainty
for claims the paper merely mentions in passing) and direction inversion (misclassifying opposing evidence as sup-
porting). Recent benchmarking confirms that state-of-the-art systems often fall short of production-level precision
on tasks requiring exhaustive retrieval and aggregation of directional claims from long documents [Liang et al., 2024].
Furthermore, because our corpus extends to 2026, LLM extraction is vulnerable to domain drift—the base models may
36

## Page 37

lack parametric knowledge of the most recent theoretical developments. As an alternative, fine-tuned models specif-
ically trained on FEP/AIF abstracts could yield higher precision than our zero-shot approach, though at a steeper
computational setup cost.
To formally bound these extraction errors and ensure robustness, a formal validation protocol is required, including a
10% manual-annotation ground-truth baseline evaluated via Cohen’s 𝜅. Such validation will verify that the automated
scoring pipeline meets minimum inter-rater reliability thresholds (𝜅> 0.70) before aggregating the data. The explicit
“irrelevant” filtering predicate further mitigates over-extraction, converting what would be a vulnerability of automated
reviews into a calibrated evidential ledger.
5.3
Research Agenda: Four Priority Next Steps
The current prototype establishes a reproducible baseline and surfaces the field’s evidence structure at corpus scale.
Four concrete next steps, ordered by the dependency chain each one unlocks, define the path from prototype to
production-grade living review.
5.3.1
Next Step 1 — Expand the Scope of Referenced Data
The present corpus of 𝑁= 819 papers is retrieved via keyword queries against three APIs (Semantic Scholar, OpenAlex,
arXiv). Three expansion axes would materially change the evidence landscape.
Additional sources. PubMed, PsycINFO, and IEEE Xplore each index Active Inference literature that the current
APIs do not reach: neuroscience clinical trials (PubMed), cognitive-behavioral studies (PsycINFO), and robotics con-
trol architectures (IEEE). For each new source, the retrieval layer requires only a source-specific connector implement-
ing the same fetch_papers(query, max_results) interface used by existing adapters. Gray literature—technical
reports, theses, and institutional preprints not yet indexed by major APIs—represents an additional tier: harvesting
from ORCID work records and institutional repositories would capture practitioner findings that never appear in
indexed venues.
Broader query coverage. The current query set is derived from the eight hypothesis keywords and their immediate
synonyms. Expanding to a full ontological synonym set (e.g., mapping “variational inference,” “surprise minimization,”
and “Helmholtz machine” as equivalent retrieval terms for FEP-related claims) would reduce the retrieval false-negative
rate for papers that use non-canonical vocabulary. A systematic evaluation of retrieval precision and recall against a
hand-curated gold-standard set of 100 known AIF papers would quantify the gap.
Custom curated bibliographies.
Domain experts can contribute citation lists directly to the corpus without
modifying any code: placing a .bib or .ris file in data/custom_bibliographies/ triggers the deduplication merge
on the next pipeline run. This pathway is the lowest-friction route to extending scope for researchers who maintain
personal reference libraries.
5.3.2
Next Step 2 — Extract and Verify Evidence Supporting Claims in Each Paper
The current extraction pipeline operates exclusively on abstracts. Abstracts contain the claims authors choose to
foreground, not necessarily the claims best supported by the paper’s data. Three mechanisms bridge this gap.
Full-text ingestion. For the subset of papers with open-access PDFs (approximately 60–70% of recent AIF preprints
on arXiv), Stage 3 can be extended to parse full-text sections—specifically Methods, Results, and Discussion—using
a structured chunking strategy that splits documents into ~512-token segments aligned to section boundaries. The
existing nanopublication schema accommodates a source_section field (currently unused) that would record the
provenance of each extracted assertion (abstract vs. results vs. discussion), enabling downstream stratification of
evidence by rhetorical function.
Claim-evidence pairing.
The current extraction prompt asks the LLM to classify a paper’s stance toward a
hypothesis but does not require it to quote the specific sentence or data point that justifies the classification. A
revised prompt would require the model to (a) identify the hypothesis-relevant passage verbatim, (b) classify the
stance, and (c) rate confidence on the basis of whether the passage reports an empirical measurement, a theoretical
derivation, or an assertion without quantitative support. This three-field extraction — evidence_quote, stance,
evidence_type — upgrades the nanopublication from a classification label to a traceable evidential pointer. For H3
(Markov Blanket Realism), where the 4 contradicting assertions drive a contested score, reviewers could then inspect
the actual quoted passages rather than trusting the LLM classification in isolation.
37

## Page 38

Human spot-check coverage. The planned 10% manual-annotation baseline would focus on inter-rater agreement;
this manual validation has not yet been performed. Extending spot-checks to verify that the extracted evidence quote
actually appears in the source document (a verbatim-match check) adds an additional fidelity gate beyond stance
accuracy.
5.3.3
Next Step 3 — Tie Hypotheses to Real-World Outcomes
The eight tracked hypotheses are formulated at the level of theoretical constructs (e.g., “the FEP provides a universal
account of self-organizing systems”). Practical applicability requires mapping from hypothesis support to observable
real-world outcomes, distinguishing which claims are actionable from which remain theoretical scaffolding.
Outcome taxonomy. Each hypothesis should be annotated with a set of outcome indicators: specific, measurable
real-world results whose observation would constitute evidence for or against the hypothesis under the closest empir-
ical operationalization. For example: - H4 (Predictive Coding): outcome indicator = reduction in prediction-error
amplitude as measured by ERP N400 or oscillatory gamma-band response in human neuroimaging studies. - H5
(Scalability): outcome indicator = task performance on standard RL benchmarks (Atari, MuJoCo, ProcGen) at or
above the performance of model-free SOTA at matched computational budgets. - H6 (Clinical Utility): outcome in-
dicator = statistically significant improvement on standardized psychiatric assessment scales (PANSS, BDI-II, PTSD
Checklist) in at least one registered clinical trial. - H7 (Morphogenesis): outcome indicator = quantitative recapitu-
lation of at least one morphogenetic patterning sequence (e.g., digit formation timecourse, limb bud size scaling) in a
computational model governed by FEP dynamics rather than reaction-diffusion equations.
For each hypothesis,
the extraction pipeline can be extended to tag assertions whose evidence type is
empirical_measurement and whose outcome aligns with these indicators, producing a filtered score that counts
only outcome-linked evidence. This outcome-filtered score sits alongside the current citation-weighted score in the
hypothesis table, providing a direct answer to “how much of this support is grounded in real-world observations
rather than theoretical commentary?”
Application domain cross-walk. The subfield classification (A1–C5) already partitions the corpus by application
domain. Intersecting hypothesis scores with application domain membership—computing score(𝐻𝑖, 𝐷𝑗) for each hy-
pothesis 𝐻𝑖and domain 𝐷𝑗—would reveal which domains are generating empirical traction versus theoretical citation
counts. H1 (FEP Universality) likely has high A2 (philosophy) support and lower C1–C5 empirical support; quan-
tifying this split would replace qualitative description of the “neutral plurality” with a decomposed evidence profile
grounded in domain labels already computed by Stage 2.
5.3.4
Next Step 4 — Formal Evaluation Rubric for Pipeline Quality
The current validation is primarily structural: do scripts run, do outputs exist, do tests pass, does the PDF render?
A formal evaluation rubric answers a different question: how accurate is the evidence landscape this system produces?
Four rubric dimensions, together with their measurement protocols and target thresholds, define what “good enough
for a published living review” means.
Table 11: Proposed evaluation rubric for pipeline quality assessment. Each dimension has a measurement protocol,
a current baseline, and a target threshold for a production living review. All metrics are computed on a held-out
annotation set of 200 randomly sampled assertions.
Dimension
Protocol
Current
Target
Extraction direction accuracy
Cohen’s 𝜅(human vs. LLM stance)
not measured
𝜅> 0.80
Evidence-quote fidelity
Verbatim substring match rate
not measured
≥90%
Corpus recall
Precision/recall vs. 100-paper gold set
not measured
recall ≥0.85
Outcome grounding rate
Fraction of supporting assertions citing an outcome indicator
not measured
≥30%
The four rubric dimensions map directly to the four next steps: corpus recall measures Step 1 progress, evidence-quote
fidelity measures Step 2 progress, outcome grounding rate measures Step 3 progress, and extraction direction accuracy
is the targeted baseline to be established as the other three improve. Reporting all four numbers alongside hypothesis
scores in each pipeline release converts a qualitative description of limitations into a versioned, trackable quality
scorecard. This transforms the current “we acknowledge limitations” posture into an audit trail: readers can see
whether the rubric scores improved between release v1.0 and v2.0, and reviewers can evaluate pipeline trustworthiness
on principled criteria rather than subjective judgement.
38

## Page 39

5.4
Future Directions: Beyond Tally-Based Evidence Aggregation
Beyond the four priority next steps above, the scoring machinery itself can be upgraded. We identify four directions,
ordered by expected impact.
5.4.1
Hierarchical Bayesian Hypothesis Scoring
The most direct extension replaces the additive tally with a hierarchical Bayesian model that treats each hypothesis
score as a latent variable inferred from noisy assertion observations. Under this formulation, each assertion 𝑎𝑖con-
tributes a likelihood term 𝑃(𝑎𝑖|𝜃𝐻, 𝜎) parameterized by the hypothesis-level evidence strength 𝜃𝐻and an observation
noise term 𝜎capturing LLM extraction uncertainty. A hierarchical prior 𝜃𝐻∼𝒩(𝜇field, 𝜏2) pools information across
hypotheses, enabling principled shrinkage for hypotheses with sparse evidence (e.g., H6 Clinical Utility, with only 25
assertions). This framework produces posterior credible intervals rather than point estimates, providing uncertainty
quantification that the current tally-based scores lack. Temporal dynamics can be modeled through time-varying
parameters 𝜃𝐻(𝑡) using state-space formulations that re-weight older evidence rather than treating all cumulative
assertions equally.
5.4.2
Causal Evidence Graphs
A second-generation knowledge graph would encode not only assertion-level relationships (paper →supports →hy-
pothesis) but also causal dependencies among hypotheses themselves.
For example, evidence for predictive
coding (H4) often implicitly supports FEP universality (H1), yet the tally-based approach treats them as indepen-
dent. A causal evidence graph—structured as a directed acyclic graph (DAG) over hypotheses with edge weights
learned from co-assertion patterns—would enable cross-hypothesis evidence propagation using belief propagation or
variational message passing. This is particularly relevant for the Active Inference literature, where hypotheses are
theoretically nested: FEP universality (H1) logically entails predictive coding (H4), and Markov blanket realism (H3)
is a prerequisite for certain formulations of H1. Encoding these dependencies would prevent the double-counting of
evidence from papers that support multiple related hypotheses and enable identification of which specific claims drive
support for downstream hypotheses. The resulting causal structure itself would be a scientific contribution—a formal
map of evidential dependencies within the field’s theoretical architecture.
5.4.3
Evidential Diversity and Source Weighting
The current formula weights assertions by log(1+citations)⋅confidence, treating all assertion sources symmetrically. A
more nuanced approach would introduce an evidential diversity index that downweights correlated evidence from
papers sharing authors, institutions, or methodological approaches. Concretely, assertions could be weighted by the
inverse of their similarity to previously counted assertions, measured via cosine similarity of paper embeddings. This
would address the observation that H1 (FEP universality) accumulates a large neutral tally partly because many A2
(philosophy) papers invoke the FEP without independently testing it—a form of evidential redundancy that inflates
the evidence base without adding independent information. Additionally, assertions could be stratified by evidence
type (empirical, theoretical, review) with configurable type-specific weights, enabling users to compute evidence scores
that privilege experimental results over theoretical commentary.
5.4.4
Agentic LLM Extractors and Domain Adaptation
Drawing on recent work extending active inference into artificial reasoning [Friston et al., 2025] and proposing AIF
as a reward-free alternative for LLM-based agents [Wen, 2025], replacing static prompt templates with goal-directed
reasoning architectures could significantly improve confidence calibration. As demonstrated by Friston et al. [Friston
et al., 2025], “active reasoning” enables agents to perform structure learning—determining which causal rule governs a
situation by seeking observations that explicitly disambiguate competing hypotheses about world models. Applied to
literature extraction, analogous uncertainty-aware reasoning could treat each paper as a structured observation to be
parsed against hypothesis definitions via an optimal experimental design rubric—directly operationalizing Next Step
2’s claim-evidence pairing at scale. The framework is domain-agnostic by design; adaptation to foundation models,
quantum computing, or synthetic biology requires only domain-specific hypothesis definitions and keyword lists within
the A/B/C taxonomy. The broader convergence between AIF and deep learning demonstrated by AXIOM [Heins
et al., 2025]—which plans in object-centric state-spaces—further validates this trajectory. Systematic cross-referencing
with the Energy-Based Model research program [LeCun et al., 2006]—including Helmholtz machines [Dayan et al.,
39

## Page 40

1995], contrastive divergence training [Hinton, 2002], and variational autoencoders [Kingma and Welling, 2014]—would
illuminate shared mathematical structures currently obscured by disciplinary siloing.
5.5
Limitations
Three constraints bound the current findings. First, extraction operates on abstracts only: full-text methods, results,
and supplementary data—where quantitative effect sizes and experimental controls live—are not yet parsed. The
rubric’s evidence-quote fidelity dimension (Step 4, Table~11) will quantify exactly how much signal this omission
suppresses once a full-text pilot is run. Second, keyword-based retrieval across three APIs produces a snapshot with
systematic false negatives: papers using non-canonical terminology, gray literature, and domain-adjacent work (EBM,
Bayesian brain models) are undercounted. The corpus recall metric provides a principled bound on this gap rather
than a vague acknowledgement of it. Third, the citation-weighted tally treats all assertion sources symmetrically;
the evidential diversity and outcome-grounding extensions above are the concrete remedies. These are not general
disclaimers but tracked deficits against which the Step 1–4 roadmap makes measurable progress.
5.6
Broader Impact
Knight et al. [Knight et al., 2022] identified three capabilities as goals for the field: “encompass increased scope of
relevant works,” “integrate multiple forms of annotation and participation,” and “facilitate integration of manual
and artificial contributions.”
The four-step research agenda in §5.3 operationalizes each of these directly: Step 1
addresses scope, Step 2 addresses the quality of extracted contributions, Step 3 addresses empirical grounding, and
Step 4 provides the formal rubric that makes “integration of manual and artificial contributions” verifiable rather than
aspirational.
By demonstrating that LLM-driven assertion extraction can produce scalable, queryable representations of scientific
evidence—processing 𝑁= 819 papers spanning approximately two and a half decades (2005–2026), extracting struc-
tured assertions, and evaluating 8 core hypotheses—this work provides a reusable architecture for realizing this vision.
The corpus window begins in 2005 to capture Energy-Based Model and variational Bayesian antecedents that predate
the Free Energy Principle label itself; the formal FEP was introduced in 2006 [Friston et al., 2006] and reached its core
elaboration by 2010 [Friston, 2010]. The citation network metrics (2,176 edges, 0.33% density, mean in-degree 2.7)
characterize the field’s structure, which has grown at a 20.36% CAGR while diversifying across 5 application domains.
The limitations of keyword-based retrieval across disjoint academic repositories mean that any retrieved corpus will
contain both false positives and false negatives. There is no single threshold that perfectly defines inclusion or exclusion
for a dynamic, interdisciplinary research field. The primary contribution of this work is therefore not a definitive corpus
but an open-source, modularly updatable, and versioned software package. This tool is built in reference to custom
literature bibliographies that can be iteratively curated for relevance by the community.
The combination of multi-source retrieval, LLM-based extraction, and probabilistic knowledge graph construction
provides a reusable template that advances each of these goals.
A complementary pathway is emerging through
Retrieval-Augmented Generation (RAG) architectures that ground LLMs directly in knowledge graphs, reducing
hallucination and enabling real-time, context-aware reasoning over structured evidence [Fan et al., 2024]. Integrating
our nanopublication graph into such a RAG system would enable natural-language querying of the evidence base,
further lowering the barrier for community engagement. The recent release of nanopub-js v0.1.0 [Knowledge Pixels,
2026]—enabling browser-based creation, signing, and querying of nanopublications—lowers the barrier for community-
contributed assertions, bringing the participatory evidence curation envisioned by Knight et al. within practical reach.
As LLM capabilities improve and standardized metadata adoption grows, the cost of maintaining such systems will
decrease while their utility increases. By open-sourcing the pipeline and publishing the schema, we provide both a
concrete tool for the Active Inference community and a modular blueprint that other fields can adapt and refine.
Data and code availability. The pipeline source code, configuration, and manuscript templates are available in the
project repository (see metadata.repository in config.yaml or the manuscript front matter). Nanopublications are
persisted as JSON Lines (for incremental runs) and RDF/TriG (nanopub.net-compliant); both can be archived with
the code release or on a data repository (e.g., Zenodo) for citation and long-term access.
Community recommendations, actionable implications, and open questions arising from this work are detailed in the
Discussion.
40

## Page 41

6
Discussion: Implications and Community Recommendations
6.1
Relationship to Prior Development Directions
Knight, Cordes, and Friedman [Knight et al., 2022] identified six development directions for systematic Active Inference
literature analysis: (1) increased scope of relevant works, (2) richer annotation schemes, (3) integration of manual
and artificial contributions, (4) transferable approaches across fields, (5) participation by diverse contributors, and
(6) updated analyses tracking the field’s evolution. This pipeline directly addresses directions 1, 2, 3, and 6: it scales
retrieval to three databases, replaces manual annotation with LLM-driven extraction while preserving human review
pathways, and produces a pipeline designed for incremental re-execution as new literature appears. Directions 4 and
5—cross-field transferability and community participation—remain open and are addressed below.
6.2
Tactical and Strategic Priorities
6.2.1
Adopt Rigorous Reporting Metadata
Papers should systematically report DOIs, ORCIDs, and explicit hypothesis commitments. Submitted preprints should
forward-link to their published versions to prevent fragmented citation subgraphs. Our extraction pipeline prioritizes
the DOI as the canonical identifier; failing that, deduplication cascades to arXiv IDs, Semantic Scholar IDs, and
OpenAlex IDs. Broad DOI adoption would resolve the cross-source mismatch problem, enabling higher-resolution
evidence mapping.
6.2.2
Explore Open Knowledge Graph Infrastructure
We encourage the exploration of federated nanopublication server architectures to house community-contributed as-
sertions. This would enable a continuously updated living literature review that incorporates new findings as they
are published. The release of nanopub-js v0.1.0 [Knowledge Pixels, 2026] makes browser-based creation and querying
of nanopublications practical, enabling researchers to contribute assertions directly from web interfaces. Integrating
this approach with the Active Inference Institute’s Knowledge-Engineering infrastructure [Knight et al., 2022] could
provide the standardized semantic vocabulary necessary for rigorous cross-study comparison.
6.2.3
Standardize the Ontological Lexicon
Immediate future extraction cycles should align assertion predicates with the formally curated Active Inference Ontol-
ogy. Enforcing shared ontological primitives across studies will accelerate the aggregation of evidence from otherwise
siloed research communities, advancing the interoperability goal outlined by Knight et al. [Knight et al., 2022].
6.3
Empirical and Theoretical Imperatives
6.3.1
Architect Unified Performance Benchmarks
The computational tools domain (B) lacks standardized performance benchmarks for direct comparison against deep
reinforcement learning architectures.
Establishing baseline metrics analogous to standard RL environments (e.g.,
OpenAI Gym) is a prerequisite for transitioning theoretical proposals into applied systems.
6.3.2
Prioritize Empirical Validation
Biology (C5) and Language (C3) have established theoretical frameworks but limited empirical validation. Targeted
experiments designed to test specific FEP-derived predictions—such as demonstrating morphogenesis as Bayesian
inference or measuring active inference advantages in language tasks—would strengthen the evidence base beyond
what further theoretical work alone can achieve.
6.4
Living Review Maintenance
The pipeline is designed for continuous operation rather than one-time analysis.
Incremental resume capabilities
(checkpoint-based assertion extraction, merge-on-add corpus deduplication) enable periodic re-execution as new papers
are indexed. We envision a maintenance cycle in which the pipeline is re-run quarterly, with updated hypothesis
scores and field statistics published alongside the pipeline release. Community contributors can extend the framework
by adding custom hypothesis definitions, alternative keyword taxonomies, or domain-specific extraction prompts—all
configurable via the YAML configuration file without modifying source code. A complementary long-term trajectory is
41

## Page 42

toward RAG-enabled access: integrating the nanopublication knowledge graph into a Retrieval-Augmented Generation
architecture [Fan et al., 2024] would enable natural-language querying of the evidence base, making quantitative
literature synthesis accessible to researchers without programming expertise.
6.4.1
Agentic Workspaces and MCP Integration
Beyond traditional open-source maintenance, the repository is architected as an intrinsically agentic workspace. Every
underlying source module (e.g., src/knowledge_graph/, src/visualization/) is governed by dedicated SKILL.md
files serving as Model Context Protocol (MCP) prompt-boundaries. These explicitly define the “rules of engagement”
for autonomous AI inference agents—such as enforcing the zero-mock testing philosophy via local HTTP proxies,
handling specific LLM fallback parsing logic, and respecting headless rendering constraints. This design ensures that
future AI orchestrators can natively interface with, scale, and refine the computational meta-analysis pipeline safely
and deterministically without structural micromanagement.
6.4.2
The Discovery Engine and Future Architectures
Broadening our synthesis of knowledge graphs and LLMs, future iterations of this pipeline may interface with ar-
chitectures like the Discovery Engine [Baulin et al., 2025]. This comprehensive framework is designed to overcome
the limitations of the document-centric publishing paradigm by transforming unstructured scientific literature into a
machine-operable “world model.” Their approach uses systematic, self-consistent LLM distillation to extract typed
“knowledge artifacts” from publications, which are sequentially assembled into a hierarchical Conceptual Nexus Model
(CNM) graph and encoded as a high-dimensional Conceptual Nexus Tensor. By explicitly modeling experimental
variables, causal relations, and evidential contradictions within a FAIR-aligned representation, this architecture en-
ables AI agents to mathematically navigate the knowledge landscape, trace provenance, and generate novel hypotheses
through operations akin to tensor factorization and Vector Symbolic Architectures (VSA). This shift from static dig-
ital libraries to a computable, relation-rich evidence graph deeply parallels our objective of translating unstructured
Active Inference literature into a quantifiable assertion tracking system.
6.5
Open Questions
This meta-analysis surfaces four empirically testable questions whose answers would directly advance the four-step
research agenda outlined in §5.3.
• Recency bias in citation weighting (Methodological limitation). The citation-weighting function 𝑤(𝑎) =
log(1 + citations) ⋅confidence systematically underweights recent papers (2024–2026) which have few citations.
A 2024 paper with 1 citation is weighted approximately 0.69×versus a 2015 paper with 100 citations at approx-
imately 4.6×. Future work may explore time-decay normalization to mitigate this recency penalty.
• Domain classifier over-assignment to A2 (Philosophy). The keyword-based domain classifier tends to
over-assign papers to the broad A2 (philosophy) category, where FEP universality is implicitly invoked but rarely
explicitly tested. This classification bias likely inflates H1’s neutral evidence count and should be addressed in
future work through embedding-based classification or expert annotation.
• Classifier calibration (feeds Step 1). What proportion of A1 (Formal Theory) papers would be reclassified
under an embedding-based or expert-annotated scheme, and how does this affect the field’s theoretical core? An
embedding-classifier trained on a 200-paper labeled set and evaluated on held-out A1 vs. A2 examples would
quantify the fraction of “philosophy” papers that carry formal mathematical content, directly sharpening both
retrieval scope and outcome-grounding rate.
• Falsifiability and explicit testing (feeds Step 3). H1 (FEP Universality) produces a predominantly neutral
evidence profile, consistent with the critique that FEP accommodates any behavior without generating distinc-
tive predictions [Colombo and Seriès, 2021]. Can hypothesis definitions—and author reporting standards—be
reformulated to require a formal, refutable empirical prediction before contributing a supporting assertion? The
proposed outcome-indicator taxonomy (§5.3) would operationalize this: only assertions paired with a measurable
outcome indicator would count as empirical support, converting the neutral H1 tally into a decomposed “invoked
vs. tested” breakdown.
• The Scalability Gap (feeds Step 3). H5 (AIF Scalability) shows a strong positive trend, yet head-to-head
comparisons with deep RL remain concentrated on a narrow set of benchmarks (predominantly low-dimensional
discrete environments).
Beyond what state-space dimensionality and reward density does the expected-free-
energy exploration advantage of model-based AIF degrade relative to model-free architectures such as SAC or
42

## Page 43

PPO? Answering this requires assembling the outcome-indicator-tagged evidence (Step 3) and identifying which
benchmark comparisons are already in the literature versus which are genuinely absent.
• Evidence Cross-Pollination (feeds Step 1 + Step 4). To what extent do mathematical structures un-
derlying variational free energy minimization and energy function optimization in Energy-Based Models (VAEs,
contrastive divergence) converge? Extending the corpus to include EBM literature (Step 1) and running the
assertion extractor on the merged set would produce a cross-domain hypothesis score for the shared-architecture
claim—a direct test of convergence rather than a theoretical argument. The rubric’s corpus recall metric (Step
4) would validate whether the expanded retrieval actually captures the EBM literature at recall ≥0.85.
6.6
Pipeline as a Community Instrument
The four next steps are not a private development roadmap—they are an invitation. The repository is structured so
that each step can be contributed incrementally: a new source connector (Step 1), a revised extraction prompt with
evidence-quote fields (Step 2), a YAML file defining outcome indicators per hypothesis (Step 3), and an annotation
script that computes rubric scores against a provided gold set (Step 4). None of these require modifying the scoring
engine or the knowledge graph schema. By publishing the rubric thresholds alongside the current baseline scores,
this work makes explicit what it would take for a community contributor to demonstrably improve the system—and
provides the tooling to verify that improvement without relying on subjective assessment.
6.7
Limitations
Recency bias: The citation-weighting function 𝑤(𝑎) = log(1+citations)⋅confidence systematically underweights recent
papers (2024–2026) which have few citations. A 2024 paper with 1 citation is weighted ∼0.69× versus a 2015 paper
with 100 citations at ∼4.6×. Future work may explore time-decay normalization.
Classifier bias: The assertion counts are also sensitive to corpus composition: H1’s large neutral tally (429) partially
reflects the keyword classifier’s tendency to assign papers to the broad A2 (philosophy) category, where FEP univer-
sality is implicitly invoked but rarely explicitly tested. This classifier bias likely inflates H1’s neutral classification
count and should be addressed in future work.
43

## Page 44

7
Appendix: Tooling and Infrastructure
The practical utility of a computational meta-analysis depends on robust tooling at each pipeline stage: assertion
extraction, modeling and simulation, knowledge-graph infrastructure, and quality assurance. This appendix surveys
the open-source ecosystem of Active Inference (AIF) and Free Energy Principle (FEP) implementations as of early
2026, documents the engineering trade-offs behind our knowledge-graph backend, and lists the multi-level quality gates
enforced by the pipeline.
7.1
LLM-Based Assertion Extraction
Extracting structured assertions from unstructured text is the most labor-intensive component of knowledge-graph
construction. Manual annotation produces high-quality results but does not scale to corpora of thousands of papers—a
constraint demonstrated by Knight et al. [Knight et al., 2022], whose systematic analysis of FEP and Active Inference
publications required manual coding of structural, visual, and mathematical features for hundreds of annotated papers.
We implement a hybrid approach: an LLM performs initial extraction and human review provides validation pathways.
Our extraction pipeline deploys a locally hosted LLM through Ollama [Ollama Team, 2024]. Each paper’s abstract
is assessed against the eight hypothesis definitions in a structured prompt requesting a JSON array of assessments.
Unlike keyword matching, which detects only topical terms, the LLM evaluates the semantic relationship between a
paper’s claims and each hypothesis. Papers critiquing the FEP correctly receive “contradicts” assessments for FEP
Universality (H1), while methodology tutorials receive “neutral” assessments reflecting their pedagogical character.
Detailed prompt engineering, schemas, and failure modes are documented in the extraction pipeline section.
7.2
Software Ecosystem
The Active Inference community has developed a rapidly growing ecosystem of open-source tools spanning multiple
programming languages, inference paradigms, and application domains. This section provides a comprehensive survey
of publicly available implementations as of early 2026, organized by functional category. We emphasize tools with
accessible source code: open-source availability is a prerequisite for reproducibility and community-driven validation.
7.2.1
General-Purpose Frameworks
Six general-purpose frameworks dominate the landscape, collectively covering discrete, continuous, and real-time
inference:
pymdp. The pymdp library [Heins et al., 2022] provides a Python implementation of active inference for discrete
state-space POMDPs, supporting message passing on factor graphs, policy inference via expected free energy, and
hierarchical generative models. It has become the standard entry point for algorithm development and the most widely
forked AIF repository.
SPM. The SPM package (Wellcome Centre for Human Neuroimaging) includes MATLAB implementations of Dy-
namic Causal Modeling and variational Bayesian inference under the FEP. It remains the reference implementation
for neuroimaging applications and houses the original Friston-group POMDP scripts.
RxInfer.jl. RxInfer is a Julia package for reactive message-passing-based Bayesian inference, supporting real-time and
streaming inference suitable for robotics and online learning. Version 4.0.0 (early 2025) [Bagaev et al., 2025] introduced
projected constraints and adaptive inference optimized for dynamic data streams and autonomous systems. The RxIn-
fer ecosystem includes extensive tutorials covering Bayesian linear regression, hidden Markov models, Kalman filtering,
Gaussian process regression, hierarchical Gaussian filters, nonlinear sensor fusion, and active inference mountain car
control, available at the official documentation and the Learnable Loop tutorial portal.
ActiveInference.jl. In parallel to RxInfer’s generalized message-passing focus, ActiveInference.jl provides a Julia-
native, near drop-in conceptual analogue to Python’s pymdp [Nehrer et al., 2025]. It explicitly targets computational
psychiatry and cognitive neuroscience workflows emphasizing standard discrete-state POMDP simulation, parameter
estimation, and recovery. The library leverages Julia’s array semantics—utilizing vectors of arrays to efficiently encode
multimodal factorized models via the canonical A, B, C, D, E components—to streamline tasks such as generating
synthetic behavioral data, fitting models to subject behavior, and probing internal beliefs via robust simulation loops
(infer_states!, infer_policies!, sample_action!).
Cpp-AIF. The Cpp-AIF header-only C++ library [Gregoretti, 2023] implements active inference for discrete POMDPs
with multicore parallelization of the most demanding computational kernels—multidimensional inner products for
44

## Page 45

expected free energy computation and state estimation. By abstracting the mathematical details behind a high-level
API, Cpp-AIF targets embedded systems and performance-critical applications where Python overhead is prohibitive.
FEPS. Free Energy Projective Simulation [Pazem et al., 2024] combines active inference with interpretable graphical
policy representations, enabling agents to plan via expected free energy while exposing decision logic as human-
readable policy graphs. FEPS targets interpretable reinforcement learning tasks where black-box deep agents are
undesirable—behavioral biology, clinical decision support, and safety-critical robotics.
7.2.2
Deep Active Inference
Scaling active inference beyond tabular POMDPs to high-dimensional observation spaces requires neural-network
function approximators. A growing body of deep active inference implementations explores this direction:
The foundational deep AIF agent of Fountas et al. [Fountas et al., 2020] introduced Monte-Carlo tree search over learned
latent spaces, achieving non-trivial Atari performance. Millidge’s DeepActiveInference extended this to continuous
control with backpropagation-based world models [Millidge, 2020].
Champion’s Branching-Time Active Inference
(BTAI_3MF) and its deep variant (Deep_BTAI_3MF) implement tree-structured planning under the free-energy
objective, scaling active inference to partially observable environments with multi-step lookahead [Champion et al.,
2021]. Most recently, AXIOM [Heins et al., 2025] achieves competitive Gameworld-10k benchmark performance using
expanding object-centric world models, learning in minutes rather than hours—a landmark result for scalability.
7.2.3
Predictive Coding and Neural Generative Coding
Predictive coding provides the core computational mechanism linking active inference to neuroscience. Several imple-
mentations offer accessible entry points:
ngc-learn. The Neural Generative Coding library (ngc-learn v3.0, JAX-based) provides a framework for simulating
neurobiologically-plausible systems using predictive-coding circuits, Hebbian learning, and spike-based dynamics. It
supports constructing arbitrary neural generative models without backpropagation, directly instantiating the FEP’s
prediction-error minimization at the circuit level.
Active Neural Generative Coding (ANGC). ANGC implements a form of active inference using paired predictive-
coding circuits—an actor/policy circuit and a world/transition model—that co-evolve across episodes without back-
propagation. The agent decomposes behavior into epistemic foraging (uncertainty reduction) and instrumental (reward-
seeking) terms, operating with sparse rewards where classical DQN requires dense reward engineering.
Predictive Coding ≈Backprop. Millidge et al. demonstrate that predictive-coding networks can approximate back-
propagation along arbitrary computational graphs [Millidge et al., 2022], providing a biologically plausible alternative
to gradient descent. The PredictiveCodingBackprop repository provides the reference implementation.
7.2.4
Benchmarking Progress
The scalability gap between AIF and deep reinforcement learning has been a central limitation of the tools domain.
Recent work demonstrates significant progress on two fronts. First, AXIOM [Heins et al., 2025] outperforms state-of-
the-art model-based deep RL agents including DreamerV3 on the Gameworld-10k benchmark while using substantially
smaller model sizes; its object-centric scene decomposition enables sample-efficient learning from structured represen-
tations rather than raw-pixel memorization. Second, variational message-passing formulations [Champion et al., 2021]
connect EFE decomposition—into risk, ambiguity, epistemic (information-seeking), and instrumental (goal-reaching)
components—to practical planning algorithms, advancing the theoretical justification for EFE-based policy selection
(H2). Separately, Friston et al. [Friston et al., 2025] introduce structure learning via Bayesian Model Reduction as a
principled approach to artificial reasoning under active inference.
7.2.5
Comprehensive Open-Source Tool Survey
The following table catalogs the principal open-source Active Inference implementations surveyed, organized by func-
tional category. For each tool we list the primary language, application domain, and associated publication or repos-
itory. The table is intended as a navigational resource for researchers seeking existing implementations relevant to
specific hypotheses (H1–H8) or application domains (A1–C5).
45

## Page 46

Table 12: Comprehensive open-source survey of Active Inference and Free Energy Principle software, grouped by
functional category. Forty-plus implementations span seven categories.
Tool / Repository
Lang.
Description
Paper / Source
General-Purpose Frameworks
pymdp
Python
Discrete POMDP active inference;
factor
graphs, hierarchical models
Heins et al. [2022]
SPM
MATLAB
DCM, variational Bayes; neuroimaging refer-
ence implementation
Friston
et
al.
[2017]
RxInfer.jl
Julia
Reactive message passing; real-time stream-
ing Bayesian inference
Bagaev
et
al.
[2025]
ActiveInference.jl
Julia
Discrete POMDP AIF; parameter recovery
for computational psychiatry
Nehrer
et
al.
[2025]
Cpp-AIF
C++
Header-only POMDP AIF library with multi-
core parallelization
Gregoretti [2023]
FEPS
Python
EFE on interpretable policy graphs; projec-
tive simulation
Pazem
et
al.
[2024]
ActivPynference
Python
Discrete AIF with factor-graph message pass-
ing; educational focus
—
pypc
Python
Predictive-coding inference engine for contin-
uous models
—
ActiveInferAnts
Rust
Rust-native AIF framework with WASM com-
pilation target
—
Deep Active Inference
deep-active-inference-
mc
Python
Monte-Carlo tree search in learned latent
spaces; Atari
Fountas
et
al.
[2020]
DeepActiveInference
Python
Continuous deep AIF with backprop-based
world models
Millidge [2020]
BTAI_3MF
Python
Branching-time
AIF
with
multi-step
tree
planning
Champion et al.
[2021]
Deep_BTAI_3MF
Python
Deep neural variant of BTAI with learned
state spaces
Champion et al.
[2021]
OO-BTAI_3MF
Python
Object-oriented BTAI variant for structured
environments
—
AXIOM
Python
Object-centric world models; Gameworld 10k
in minutes; beats DreamerV3
Heins et al. [2025]
Deep-AIF-POMDPs
Python
Deep AIF for partially observable MDPs
—
Homing-Pigeon
Python
Navigation agent using deep active inference
—
active-inference
(Voostrum)
Python
Continuous deep AIF with learned generative
models
arXiv:2406.07726
Predictive Coding & Neural Generative Coding
ngc-learn
Python/JAX Neurobiological simulation; predictive-coding
circuits, Hebbian learning
—
ANGC
Python
Backprop-free AIF agent with paired PC cir-
cuits
AAAI 2022
PredictiveCodingBackpropPython
Predictive coding approximates backprop on
arbitrary graphs
Millidge
et
al.
[2022]
Supervised-Predictive-
Coding
Python
Supervised learning via hierarchical predic-
tive coding
—
predcoding
Python
Minimal predictive-coding implementation
—
pybrid
Python
Hybrid predictive-coding and active-inference
library
—
nmpassing
Python
Neural message passing for PC networks
—
Neuroscience, Embodied & Biological
allostasis
Python
Allostatic regulation via AIF; interoceptive in-
ference
bioRxiv:2021.02.16
Continued on next page
46

## Page 47

Tool / Repository
Lang.
Description
Paper / Source
ants
Python
Ant foraging simulation with stigmergic AIF
agents
Heins et al. [2024]
Reward_Bases
Python
Reward-basis function representations under
AIF
bioRxiv:2022.04.14
action-oriented
Python
Action-oriented predictive-processing models
Tschantz
et
al.
[2020]
Biofirm
Python
Bioregional stewardship via organizational
AIF
—
bayesian-mechanics-
sdes
Python
Bayesian mechanics:
SDE simulations of
Markov-blanket dynamics
arXiv:2206.02629
reverse_engineering
MATLAB
Reverse-engineering neural dynamics under
the FEP
—
Multi-Agent & Social Dynamics
opinion_dynamics
Python
Opinion dynamics and belief formation via
AIF
—
network-actinf
Python
Network-level active inference with coupled
agents
—
Variational-Capsule-
Routing
Python
Capsule networks with variational inference
routing
AAAI 2020
Active-Inference-
Successor
Python
Successor representations under active infer-
ence
—
Domain-Specific Applications
adaptive_aif_agents_fl Python
Adaptive AIF agents for federated learning
arXiv:2410.09099
smartville
Python
IoT smart-building control via AIF under par-
tial observability
TechRxiv 2025
FEP_Blorpomon
Python
Game-theoretic AIF agent demonstration
—
MountainCarAI
Python
Mountain car control via active inference
—
rl-inference
Python
Bridging RL and active inference policy selec-
tion
arXiv:2002.12636
EFE-GLean
Python
Expected free energy with generalized learn-
ing
Entropy 2025
EFEasVFE
Julia
EFE reformulated as variational free energy
—
Robust-FE-
Minimization
Python
Robust decision-making via free-energy mini-
mization
arXiv:2503.13223
Tutorials & Educational Resources
Active-Inference-from-
Scratch
Python
Step-by-step AIF implementation tutorial
—
IC2S2-AIF-Tutorial
Python
Computational social-science AIF tutorial
—
julia4ta
tutorials
(9x10–12)
Julia
RxInfer-based AIF agent tutorials
—
ActInf Textbook Co-
lab
Python
Interactive notebooks for Parr et al. [2022]
—
deep_aif_workshop
Python
Workshop materials for deep active inference
—
AdaptiveResonance.jl
Julia
Adaptive resonance theory models in Julia
—
47

## Page 48

7.2.6
Comparative Feature Matrix
Table 13: Comparative feature matrix of seven representative Active Inference packages. Features span language,
state-space type, inference algorithm, hierarchical support, GPU acceleration, license, and primary use case.
Feature
pymdp
SPM
RxInfer.jl
ActiveInf.jl
Cpp-AIF
FEPS
ngc-learn
Language
Python
MATLAB
Julia
Julia
C++
Python
Python/JAX
State Spaces
Discrete
Disc.+Cont.
Continuous
Discrete
Discrete
Discrete
Continuous
Inference
Msg. pass.
Var. Bayes
Reactive msg.
Msg. pass.
EFE+state
EFE on graphs
Pred. coding
Deep AIF
Partial
No
Custom factors
No
No
No
Yes
Real-time
No
No
Yes
No
Yes
No
No
Hierarchical
Yes
Yes (DCM)
Yes
No
Yes
No
Yes
GPU
No
No
No
No
CPU multi
No
Yes (JAX)
License
MIT
GPL
MIT
MIT
MIT
MIT
BSD-3
Primary Use
Prototyping
Neuroimaging
Robotics
Comp. psych.
Embedded
Interp. RL
NeuroAI
The complementary strengths across these packages reflect a fragmented but maturing ecosystem. The survey reveals
several patterns: (1) Python dominates ($￿$75% of implementations), with Julia emerging as the preferred alternative
for performance-critical applications; (2) discrete-POMDP implementations outnumber continuous variants by ap-
proximately 3:1, reflecting pymdp’s community influence; (3) deep active-inference implementations are concentrated
in a small number of research groups (Champion, Millidge, Fountas, Heins), suggesting high barriers to entry; (4)
multi-agent and social AIF implementations remain sparse relative to single-agent tools; and (5) domain-specific ap-
plications (IoT, federated learning, smart buildings) represent the newest and fastest-growing category, aligning with
the temporal growth patterns observed in the C-domain (applied) subfields. The variational-free-energy foundations
shared by Active Inference and Energy-Based Models—including Helmholtz machines [Dayan et al., 1995], Boltzmann
machines [Hinton, 2002], and variational autoencoders [Kingma and Welling, 2014]—suggest that interoperability with
mainstream deep generative-modeling frameworks (PyTorch, JAX) could bridge these parallel research programs.
7.3
Knowledge Graph Infrastructure
Our knowledge graph uses an RDF-compatible schema deployable on standard semantic-web infrastructure.
The
nanopublication model [Groth et al., 2010, Kuhn et al., 2016] provides a principled atomic unit of scientific evidence:
each nanopublication packages a single assertion (e.g., “Paper X supports Hypothesis Y”) with explicit provenance
and publication metadata in four named RDF graphs (Head, Assertion, Provenance, Publication Info). This structure
satisfies the FAIR data principles by design: nanopublications are Findable via URI-based identification, Accessible
through standard RDF protocols, Interoperable via W3C-standard TriG serialization, and Reusable with explicit
provenance and CC0 licensing. The full RDF schema and a TriG serialization example are presented in the methodology
and Appendix~8.5.
The engineering trade-offs among the three deployment options are straightforward:
Nanopublication servers provide decentralized, content-addressed storage. The pipeline writes nanopublications in
two forms: JSON Lines (for incremental checkpointing and tooling) and RDF/TriG per the nanopublication standard
(Assertion, Provenance, Publication Info), suitable for the nanopublication network and FAIR deployment.
The
recent release of nanopub-js v0.1.0 [Knowledge Pixels, 2026]—a JavaScript library enabling browser-based creation,
signing, and querying of nanopublications—opens the possibility of community-contributed assertions directly from
web interfaces, lowering the barrier to participatory evidence curation. Future integration with Trusty URIs [Kuhn and
Dumontier, 2014] would provide cryptographic content verification and persistent identifiers for each nanopublication.
RDF stores (e.g., Apache Jena Fuseki, Blazegraph, Oxigraph) enable SPARQL queries such as “find all papers
supporting hypothesis 𝐻published after 2020 in the neuroscience domain (C1).” The cost is operational overhead and
query latency.
Property-graph databases (e.g., Neo4j) prioritize traversal performance for path queries and community detection,
at the expense of semantic-web compatibility.
While RDF and property graphs excel at structurally organizing assertions, it is crucial to recognize that they inherently
compress the rich epistemic context of the original papers (e.g., methodological caveats, sample sizes, scope limitations)
into flattened confidence scores—a fundamental limitation of current automated knowledge extraction discussed in
the conclusion.
48

## Page 49

The Active Inference Ontology namespace ensures integration with external ontologies and linked-data resources.
7.4
Multi-Level Quality Assurance
Quality assurance operates at four levels: assertion-level confidence and review, graph-level structural consistency,
score-level boundary tests, and pipeline-level continuous-integration coverage.
7.4.1
Assertion-Level Validation
Assertions below a configurable confidence threshold (default 0.6) are flagged for review. Inter-annotator agreement
(𝜅) is computed when multiple annotators assess the same paper. The threshold is chosen to balance recall against
the prompt-engineering cost of pushing the LLM to over-commit; lowering it inflates noisy neutral assertions, raising
it discards weakly supported but legitimate claims.
7.4.2
Graph-Level Consistency Checks
Consistency checks verify that all nodes link to valid targets and no orphan nodes exist. Coverage metrics track the
proportion of annotated papers, the fraction of references that resolve inside the corpus, and the per-domain assertion
density.
7.4.3
Score-Level Unit Testing
Hypothesis scoring is validated through unit tests on synthetic data verifying boundary conditions: all-support fixtures
must produce scores at +1, all-contradict at −1, and balanced inputs at 0. Sensitivity analysis sweeps over confidence
thresholds and citation-weighting schemes to confirm that qualitative rankings are stable.
7.4.4
Pipeline-Level Test Coverage
Test-driven development enforces 90% minimum code coverage on project modules and 60% on shared infrastructure,
with real data and computation (no mocking). All tests run on every push; failures block merges and releases.
7.4.5
Quality Thresholds
Table 14: Multi-level quality-assurance thresholds enforced across the pipeline. Each level defines a metric, minimum
threshold, and failure action. Pipeline-level thresholds (90% coverage, 100% pass rate) are enforced via CI gates;
lower-level checks emit warnings or block release as indicated.
Level
Metric
Threshold
On Failure
Assertion
Confidence 𝑐
≥0.6
Flag for review
Assertion
Inter-annotator 𝜅
≥0.70
Re-annotate
Graph
Orphan-node ratio
= 0
Reject build
Graph
Corpus coverage
≥80%
Warning
Score
Boundary tests (all-support / all-contradict / balanced)
All pass
Block release
Score
Sensitivity-sweep stability
Top-𝑘ranks unchanged
Warning
Pipeline
Project-code coverage
≥90%
Block merge
Pipeline
Infrastructure coverage
≥60%
Block merge
Pipeline
Test pass rate
100%
Block release
The hypothesis-evidence results, temporal dynamics of evidence accumulation, and assertion analysis are presented in
the hypothesis results section.
49

## Page 50

8
Appendix: Mathematical and Algorithmic Details
This appendix collects the formal mathematical definitions, derivations, and algorithmic specifications referenced from
the main methodology section. Each subsection is self-contained; equations are labelled for cross-referencing from the
body and from §18.
8.1
Citation-Weighted Hypothesis Scoring Formula
For each hypothesis 𝐻, we compute a citation-weighted evidence score aggregating all assertions relevant to 𝐻:
score(𝐻) =
∑𝑎∈𝑆(𝐻) 𝑤(𝑎) −∑𝑎∈𝐶(𝐻) 𝑤(𝑎)
∑𝑎∈𝐴(𝐻) 𝑤(𝑎)
(3)
where 𝑆(𝐻) is the set of supporting assertions, 𝐶(𝐻) the set of contradicting assertions, 𝐴(𝐻) all assertions for 𝐻
(including neutral), and the weight function is
𝑤(𝑎) = log(1 + citations(𝑎)) ⋅confidence(𝑎).
(4)
The logarithmic citation weighting ensures that highly cited papers carry more influence while preventing any single
blockbuster paper from dominating the score. The score lies in [−1, 1]: values near +1 indicate strong supporting
evidence, values near −1 strong contradicting evidence, and values near 0 balanced or insufficient evidence.
As
emphasized in the main text, this score is a relative evidentiary ranking within the current literature topology, not a
calibrated Bayesian probability of the hypothesis being true.
Temporal aggregation. We additionally compute temporal trends by evaluating the cumulative score at each year
𝑡, using only assertions from papers published in year ≤𝑡:
score(𝐻, 𝑡) =
∑𝑎∈𝑆(𝐻,𝑡) 𝑤(𝑎) −∑𝑎∈𝐶(𝐻,𝑡) 𝑤(𝑎)
∑𝑎∈𝐴(𝐻,𝑡) 𝑤(𝑎)
.
(5)
This reveals whether support for a hypothesis is growing, declining, or plateauing over time. Cumulative aggregation
(rather than yearly windows) is preferred because per-year assertion counts for narrow hypotheses are too sparse for
stable point estimates.
Algorithmic specification. The scoring routine is a pure function of the assertion set; it has no hidden state and is de-
terministic given the input. The reference implementation lives in projects/act_inf_metaanalysis/src/scoring/citation_wei
function score(H, assertions):
S, C, A_all = 0, 0, 0
for a in assertions where a.hypothesis == H:
w = log(1 + a.citations) * a.confidence
if a.direction == "supports":
S
+= w
elif a.direction == "contradicts": C
+= w
A_all += w
return (S - C) / A_all
if A_all > 0
else 0
Boundary tests in tests/test_scoring.py confirm that all-support fixtures yield +1, all-contradict fixtures yield −1,
and balanced fixtures yield 0 within numerical tolerance.
8.2
Non-negative Matrix Factorization (NMF) for Topic Modeling
We apply NMF to the TF-IDF matrix of the corpus to discover latent topics.
Given the document-term matrix
𝑉∈ℝ𝑛×𝑚
≥0
, NMF finds factor matrices 𝑊∈ℝ𝑛×𝑘
≥0
and 𝐻∈ℝ𝑘×𝑚
≥0
such that 𝑉≈𝑊𝐻, where 𝑘is the number of topics.
We use multiplicative update rules [Lee and Seung, 1999]:
𝐻←𝐻⊙
𝑊𝑇𝑉
𝑊𝑇𝑊𝐻+ 𝜖,
𝑊←𝑊⊙
𝑉𝐻𝑇
𝑊𝐻𝐻𝑇+ 𝜖,
(6)
50

## Page 51

with 𝜖= 10−10 for numerical stability and a fixed random seed of 42 for reproducibility (deterministic topic alignment
across pipeline runs, with empirical stability confirmed via Jaccard similarities > 0.90 across alternative seeds).
Term-Frequency Inverse Document Frequency (TF-IDF). The document-term matrix is constructed using a
smoothed TF-IDF weighting [Salton et al., 1975]. For term 𝑡in document 𝑑:
TF-IDF(𝑡, 𝑑) = tf(𝑡, 𝑑) ⋅[log(
𝑁
df(𝑡) + 1) + 1] ,
(7)
where tf(𝑡, 𝑑) = count(𝑡, 𝑑)/|𝑑| is the normalized term frequency, 𝑁the total number of documents, and df(𝑡) the
document frequency of term 𝑡. The +1 additive smoothing in the denominator prevents division by zero and reduces
the weight of extremely rare terms; the outer +1 ensures strictly positive IDF values. Document vectors are L2-
normalized before NMF factorization.
8.3
Field Growth-Rate Estimation
The mean year-over-year growth rate
̄𝑔is the arithmetic mean of annual growth rates computed only for years
where the prior year had non-zero publications:
̄𝑔=
1
|𝑌| ∑
𝑦∈𝑌
𝑛𝑦−𝑛𝑦−1
𝑛𝑦−1
,
(8)
where 𝑌= {𝑦∶𝑛𝑦−1 > 0} and 𝑛𝑦is the number of publications in year 𝑦.
The doubling time 𝑡𝑑is derived from the mean annual growth rate:
𝑡𝑑=
ln 2
ln(1 +
̄𝑔).
(9)
The compound annual growth rate (CAGR) over the full span [𝑦0, 𝑦𝑇] is
CAGR = (𝑛cumulative(𝑦𝑇)
𝑛cumulative(𝑦0) )
1/(𝑦𝑇−𝑦0)
−1.
(10)
For the current corpus, CAGR = 20.36%. The more recent growth phase (2010–2026) exhibits substantially higher
annualized growth than the long-run average; reporting both the 𝑇-year CAGR and recent-phase CAGR avoids
overstating maturity-era expansion.
8.4
Advanced Visualization Methods
8.4.1
PCA of TF-IDF Embeddings
Principal Component Analysis (PCA) is applied to the TF-IDF matrix 𝑉to project each document into a 2-D space.
The projection preserves the directions of maximum variance, enabling visual inspection of document clustering by
domain. Loading arrows overlay the top-variance terms onto the scatter plot, showing which vocabulary drives the
principal components.
8.4.2
Hierarchical Clustering Dendrogram
For each domain 𝑠, we compute the centroid
̄𝑣𝑠=
1
|𝐷𝑠| ∑𝑑∈𝐷𝑠𝑣𝑑where 𝐷𝑠is the set of documents in domain 𝑠and
𝑣𝑑is the TF-IDF vector of document 𝑑. Ward linkage is applied to the centroid matrix to produce a hierarchical
clustering dendrogram showing semantic proximity between domains.
8.4.3
Term Heatmap
For each domain 𝑠and term 𝑡, we compute the mean TF-IDF weight
̄𝑤𝑠,𝑡=
1
|𝐷𝑠| ∑𝑑∈𝐷𝑠TF-IDF(𝑡, 𝑑). The heatmap
displays
̄𝑤𝑠,𝑡for the top-𝑘terms (by global document frequency) across all domains, with cell intensity proportional
to mean weight. This reveals distinctive vocabulary patterns that differentiate domains beyond the keyword-level
classification used for subfield assignment.
51

## Page 52

8.4.4
Term Co-occurrence Matrix
The co-occurrence matrix 𝐶∈ℝ𝑘×𝑘counts the number of documents in which two terms appear together. For top-𝑘
terms by document frequency, 𝐶𝑖𝑗= |{𝑑∶𝑡𝑖∈𝑑∧𝑡𝑗∈𝑑}|. The matrix is normalized to [0, 1] by dividing by the
maximum entry and visualized as a symmetric heatmap.
8.5
Nanopublication RDF Schema
Each nanopublication is serialized to RDF/TriG per the nanopublication standard [Groth et al., 2010, Kuhn et al.,
2016], producing four named graphs. The following annotated example illustrates the structure for a single assertion:
@prefix np:
<http://www.nanopub.org/nschema#> .
@prefix prov: <http://www.w3.org/ns/prov#> .
@prefix dc:
<http://purl.org/dc/terms/> .
@prefix aif:
<http://activeinference.institute/ontology/> .
@prefix xsd:
<http://www.w3.org/2001/XMLSchema#> .
# HEAD GRAPH: links nanopub to its three component graphs
<http://activeinference.institute/nanopub/a1b2c3d4e5f6#head> {
<http://activeinference.institute/nanopub/a1b2c3d4e5f6>
a np:Nanopublication ;
np:hasAssertion
<...#assertion> ;
np:hasProvenance
<...#provenance> ;
np:hasPublicationInfo <...#pubinfo> .
}
# ASSERTION GRAPH: the core scientific claim
<http://activeinference.institute/nanopub/a1b2c3d4e5f6#assertion> {
aif:paper/10.1038_nrn2787 aif:asserts aif:assertion/a1b2c3 .
aif:assertion/a1b2c3
aif:supports
aif:hypothesis/fep_universality ;
aif:claim
"The paper provides foundational support for FEP as a
unified brain theory."^^xsd:string ;
aif:confidence
"0.85"^^xsd:double ;
aif:citationCount
"5000"^^xsd:integer .
}
# PROVENANCE GRAPH: extraction lineage
<http://activeinference.institute/nanopub/a1b2c3d4e5f6#provenance> {
aif:assertion/a1b2c3
prov:wasGeneratedBy
<http://activeinference.institute/nanopub/a1b2c3d4e5f6> ;
prov:generatedAtTime
"2026-01-15T12:00:00+00:00"^^xsd:dateTime ;
prov:wasAttributedTo
"act_inf_metaanalysis/gemma3:4b"^^xsd:string ;
prov:hadPrimarySource aif:paper/10.1038_nrn2787 .
}
# PUBLICATION INFO GRAPH: nanopublication metadata
<http://activeinference.institute/nanopub/a1b2c3d4e5f6#pubinfo> {
<http://activeinference.institute/nanopub/a1b2c3d4e5f6>
dc:created "2026-01-15T12:00:00+00:00"^^xsd:dateTime ;
dc:creator "act_inf_metaanalysis/gemma3:4b"^^xsd:string ;
dc:license <https://creativecommons.org/publicdomain/zero/1.0/> .
}
52

## Page 53

8.5.1
Namespace Definitions
Table 15: RDF namespace definitions used in the knowledge graph and nanopublication serialization. Each prefix
maps to a W3C or domain-specific URI.
Prefix
URI
Purpose
np:
http://www.nanopub.org/nschema#
Nanopub structural predicates
prov:
http://www.w3.org/ns/prov#
PROV-O provenance model
dc:
http://purl.org/dc/terms/
Dublin Core metadata
aif:
http://activeinference.institute/ontology/
Domain-specific predicates
xsd:
http://www.w3.org/2001/XMLSchema#
XML Schema datatypes
8.5.2
Core Triple Patterns
The knowledge graph encodes five fundamental relationships:
Table 16: Core RDF triple patterns encoding the five fundamental relationships in the knowledge graph. Each pattern
links paper, assertion, hypothesis, or subfield nodes.
Triple Pattern
Meaning
Paper --aif:asserts--> Assertion
A paper makes a claim
Paper --aif:cites--> Paper
Intra-corpus citation link
Paper --aif:belongsTo--> Subfield
Domain classification
Assertion --aif:supports--> Hypothesis
Supporting evidence
Assertion --aif:contradicts--> Hypothesis
Contradicting evidence
53

## Page 54

9
Appendix:
Accessibility, Cognitive Ergonomics, and Participatory
Research Infrastructure
Automated meta-analysis tools operate at the intersection of computational infrastructure and human sensemaking.
The scalability gains demonstrated by the present pipeline are meaningful only if the resulting knowledge artefacts
remain cognitively accessible, ethically transparent, and open to diverse forms of participation. This appendix situates
our work within the broader landscape of research accessibility, cognitive ergonomics, decentralized science (DeSci),
and participatory infrastructure design, and concludes with a WCAG-mapped checklist that summarizes the concrete
accessibility practices implemented in the figure pipeline.
9.1
Cognitive Ergonomics of Knowledge Graphs
The knowledge-graph outputs of this pipeline—hypothesis dashboards, citation networks, temporal evidence
trajectories—impose nontrivial cognitive demands on users who must interpret multidimensional evidence landscapes.
Cognitive Load Theory [Sweller et al., 2011] establishes that information system designs which exceed working-memory
capacity produce disorientation and interpretive errors. Our visualization pipeline addresses this through progressive
disclosure (summarized dashboards linking to detailed per-hypothesis breakdowns), consistent visual grammars (a
fixed colour palette for supports/contradicts/neutral across all figures), and a minimum font-size floor of 16,pt that
satisfies low-vision accessibility guidelines. These are not cosmetic choices but functional requirements for trustworthy
scientific communication.
The ResNei (Research Neighbourhood) platform [Lumiruusu et al., 2025] provides a particularly instructive design
exemplar for the next generation of cognitive-ergonomic research tools. ResNei is an AI-augmented, neuro-informed
research environment that transforms heterogeneous scientific corpora into a living, collaborative knowledge graph
structured as modular Conceptual Nexus Models (CNMs). Where our pipeline produces a static (though periodically
updated) evidence snapshot, ResNei’s architecture foregrounds dynamic, responsive exploration through three cog-
nitive modes: longitudinal (tracking a concept’s evolution over time), latitudinal (surveying related concepts across
subfields), and relational (mapping connections between concepts). This trimodal navigation directly operational-
izes the progressive-disclosure principle, enabling users to manage cognitive complexity by choosing their depth of
engagement.
9.1.1
Action–Intention UX and Active Inference Design Principles
ResNei’s most theoretically significant contribution is its action–intention UX model, which replaces the conventional
passive, attention-maximizing feed with a framework that interprets user actions (uploading papers, highlighting
passages, opening concept maps, initiating discussions) as situated signals of research direction. Rather than deploying
opaque recommendation engines, the system uses explicit action trajectories to surface contextually appropriate tools
and views—an approach that resonates with the perception–action loop central to Active Inference itself [Parr et al.,
2022]. The design principle of “minimal system intervention, maximum research coherence’ ’ ensures that the interface
scaffolds orientation and affordances without interruptive prompts or aggressive automation.
This ethos directly
addresses the risk that AI-augmented sensemaking tools inadvertently narrow epistemic horizons through algorithmic
filtering.
9.1.2
Risk-Aware and Bias-Transparent Design
ResNei’s solution-design document is notable for its unusually explicit treatment of harms and ameliorations. It iden-
tifies exclusion, algorithmic misrepresentation, overconfidence in AI outputs, hidden inequalities, marginalization of
less-cited work, surveillance risks, cognitive overload, false comprehensiveness, and data privacy as first-class design
constraints [Lumiruusu et al., 2025]. Mitigations include deliberately inclusive UX (designing from the standpoint of
those usually excluded), systematic provenance and confidence indicators, framing all AI outputs as suggestions with
traceable bases, and configurable metrics beyond citation counts (e.g., conceptual novelty, geographic diversity, pub-
lication type). This risk model provides a concrete template for future iterations of our own pipeline, which currently
presents citation-weighted scores without UI-level confidence calibration or per-assertion provenance indicators.
9.2
FAIR Data and Decentralized Science
The pipeline’s outputs—nanopublications, knowledge-graph triples, and structured assertion records—are designed to
satisfy the FAIR principles (Findable, Accessible, Interoperable, Reusable) articulated by Wilkinson et al. [Wilkin-
son et al., 2016]. Each nanopublication carries machine-readable provenance (source paper DOI, extraction model,
54

## Page 55

confidence score, timestamp), enabling downstream consumers to evaluate evidential quality independently of our
aggregation choices. The JSON Lines and RDF/TriG serialization formats guarantee interoperability with existing
semantic-web infrastructure.
Decentralized Science (DeSci) represents a broader movement to dismantle structural barriers in scientific publishing
and funding through blockchain-based governance, tokenized intellectual property, and community-owned research
commons [Hamburg, 2022]. Our pipeline’s open-source, modular, and configuration-driven design aligns with DeSci
principles: the entire analytical workflow is reproducible from source code, hypothesis definitions and extraction
prompts are version-controlled in YAML rather than embedded in proprietary systems, and the nanopublication output
format is natively compatible with federated semantic publishing networks. ResNei’s architecture further advances this
trajectory by grounding its collaborative features in social accountability'' andreciprocity, interdependence, and
access’ ’ as explicit design values [Lumiruusu et al., 2025], directly addressing the power asymmetries that traditional
centralized publication systems perpetuate.
9.3
Participatory Research and Universal Access
The aspiration toward participatory research infrastructure—where diverse contributors can meaningfully engage with
evidence synthesis regardless of technical expertise—is a recurring theme across the projects discussed here. Bonney
et al.’s foundational work on citizen science [Bonney et al., 2009] demonstrated that non-expert participants can make
rigorous contributions to scientific knowledge production when provided with appropriate scaffolding, standardized
protocols, and feedback loops. Universal Design for Learning principles [Rose and Meyer, 2000] further emphasize
that accessibility is not a specialized accommodation but a design paradigm that improves usability for all users.
Applied to computational meta-analysis, this means designing systems where:
• Contribution pathways exist at multiple expertise levels—from correcting individual assertion labels (re-
quiring only domain knowledge) to extending extraction prompts or hypothesis definitions (requiring pipeline
familiarity);
• Transparency mechanisms make model confidence, extraction provenance, and aggregation logic visible and
interrogable by non-technical users;
• Multimodal access ensures that knowledge-graph outputs are available not only as programmatic APIs and
raw data files but as navigable visual interfaces with WCAG-compliant accessibility standards;
• Cultural and linguistic inclusivity is recognized as a structural requirement rather than a desirable addition—
our pipeline’s current English-language dominance (noted in §5 as a corpus bias) is a limitation that future
multilingual extraction capabilities must address.
The convergence of ResNei’s neuro-informed collaborative environment, DeSci’s decentralized governance models,
FAIR-data interoperability, and citizen-science participation frameworks collectively describes the emerging infras-
tructure requirements for equitable, cognitively supportive, and community-governed scientific sensemaking. These
are not peripheral concerns for computational meta-analysis but architectural prerequisites for systems that aspire to
serve as living, trusted evidence ledgers for rapidly evolving scientific fields.
9.4
Pipeline Accessibility Checklist
The following checklist summarizes the concrete accessibility practices implemented in the figure-generation and ren-
dering stages, mapped to the relevant Web Content Accessibility Guidelines (WCAG 2.1, Level AA) success criteria.
“Status’ ’ is recorded as Implemented, Partial, or Planned based on the current state of the pipeline.
55

## Page 56

Table 17: Accessibility practices implemented in the figure pipeline, mapped to WCAG 2.1 Level AA success criteria.
Status reflects the current pipeline; “Planned” items are tracked in the project issue tracker.
Practice
WCAG
Ref.
Implementation
Colorblind-safe palette
1.4.1
Wong (2011) 8-colour palette enforced in all figures
[Wong, 2011]; supports/contradicts/neutral encoded by
both hue and luminance.
Minimum font size
1.4.4
16 pt floor enforced in figure-generation script; tick labels
never fall below this threshold.
Sufficient contrast
1.4.3
Foreground/background contrast ≥4.5:1 for all text, ≥
3:1 for large headings, validated programmatically.
Non-color encodings
1.4.1
Direction encoded by both color and pattern (solid /
hatched / outlined) so that grayscale printing remains
interpretable.
Alt text and figure captions
1.1.1
Each \includegraphics is paired with a \caption that
describes the figure content, key axes, and main take-
away.
Consistent visual grammar
3.2.4
Domain colors, hypothesis ordering, and axis conventions
are fixed across all figures by a single style module.
Progressive disclosure
2.4.5
Summary dashboards link to per-hypothesis and per-
domain breakdowns; readers can choose depth of engage-
ment.
Machine-readable outputs
4.1.2
All analytic results published as JSON / JSONL along-
side PNG figures, enabling assistive-technology consump-
tion.
Provenance metadata
1.3.1
Each nanopublication carries source DOI, extraction
model, timestamp, and confidence; programmatically
queryable.
Multilingual extraction
—
Planned: current pipeline is English-only; future multi-
lingual prompts and corpus expansion are tracked as a
corpus-bias mitigation.
Per-assertion confidence UI
—
Planned: aggregate scores currently dominate the dash-
board; future iterations will surface per-assertion confi-
dence and rationale.
56

## Page 57

10
Notation, Abbreviations, and Glossary
This appendix consolidates the mathematical notation, abbreviations, hypothesis identifiers, and key terminology used
throughout the manuscript. Each table is self-contained and may be consulted independently. Cross-references in the
main text use the labels declared here.
10.1
Mathematical Symbols and Notation
The following symbols appear in the methodology, results, and technical appendices. Where a quantity is defined
formally, the relevant equation is referenced inline; otherwise the description here is the canonical definition. All
probabilities and confidences are real-valued in [0, 1], and all aggregate scores are in [−1, 1].
Table 18: Mathematical symbols and notation used throughout the manuscript. Scoring quantities are defined formally
in §3.6 and §8.1; growth metrics in §8.3; topic-modeling notation in §8.2.
Symbol
Description
𝑁
Corpus size after deduplication (total unique papers)
𝑛
Subfield paper count (papers within a single domain category)
𝑇= 𝑦𝑇−𝑦0
Time span in years (used for CAGR)
𝑦0, 𝑦𝑇
First and last years in the publication window
𝑛𝑦
Number of publications in year 𝑦
𝑤(𝑎)
Citation-weighted weight of assertion 𝑎: log(1 + citations) ⋅𝑐
score(𝐻)
Aggregate citation-weighted evidence score for hypothesis 𝐻, range [−1, 1]
score(𝐻, 𝑡)
Cumulative score for 𝐻using only assertions from papers published ≤𝑡
𝑆(𝐻), 𝐶(𝐻), 𝐴(𝐻)
Supporting / contradicting / all assertion sets for hypothesis 𝐻
𝑐
Assertion confidence reported by the LLM, range [0, 1]
𝑑
Assertion direction ∈{supports, contradicts, neutral}
𝑘
Number of latent topics in NMF factorization
𝑉∈ℝ𝑛×𝑚
≥0
TF-IDF document-term matrix (documents × terms)
𝑊∈ℝ𝑛×𝑘
≥0
NMF document-topic factor
𝐻∈ℝ𝑘×𝑚
≥0
NMF topic-term factor (overloaded notation; context disambiguates)
𝜖
Numerical-stability constant (10−10)
CAGR
Compound annual growth rate (Eq. 10)
𝑡𝑑
Publication doubling time in years (Eq. 9)
̄𝑔
Mean year-over-year growth rate (Eq. 8)
𝜅
Cohen’s kappa, inter-annotator agreement
tf(𝑡, 𝑑)
Normalized term frequency of 𝑡in document 𝑑
df(𝑡)
Document frequency of term 𝑡across the corpus
ℱ
Variational free energy
G
Expected free energy (used for policy ranking)
KL
Kullback–Leibler divergence
𝔼
Expectation operator
10.2
Abbreviations and Acronyms Used
The acronyms below appear at least once in the main text, methods, results, or appendices. Domain-specific shorthands
such as the A/B/C taxonomy categories (e.g., A1, A2, B, C1–C5) are documented inline at first use in the field overview
and the subfield analyses.
57

## Page 58

Table 19: Abbreviations and acronyms used in this manuscript, listed alphabetically. Where an acronym names a
software package or organization, the canonical reference appears in the bibliography.
Abbreviation
Definition
AIF
Active Inference
ANGC
Active Neural Generative Coding
API
Application Programming Interface
arXiv
Open-access preprint repository (arxiv.org)
BTAI
Branching-Time Active Inference
CAGR
Compound Annual Growth Rate
CC0
Creative Commons Zero (public-domain dedication)
CI
Continuous Integration
CNM
Conceptual Nexus Model (ResNei)
DCM
Dynamic Causal Modelling
DeSci
Decentralized Science
DOI
Digital Object Identifier
EBM
Energy-Based Model
EFE
Expected Free Energy
FAIR
Findable, Accessible, Interoperable, Reusable
FAIR4RS
FAIR Principles for Research Software
FEP
Free Energy Principle
FEPS
Free Energy Projective Simulation
HITS
Hyperlink-Induced Topic Search (Kleinberg)
IaC
Infrastructure as Code
JSON
JavaScript Object Notation
JSONL
JSON Lines (newline-delimited JSON)
KG
Knowledge Graph
KL
Kullback–Leibler (divergence)
LLM
Large Language Model
MBR
Bayesian Model Reduction
MCMC
Markov Chain Monte Carlo
MIT
Massachusetts Institute of Technology
NLP
Natural Language Processing
NMF
Non-negative Matrix Factorization
ORCID
Open Researcher and Contributor ID
PCA
Principal Component Analysis
PDF
Portable Document Format
POMDP
Partially Observable Markov Decision Process
PROV-O
PROV Ontology (W3C provenance data model)
RBM
Restricted Boltzmann Machine
RDF
Resource Description Framework
ResNei
Research Neighbourhood (cognitive-ergonomic platform)
RL
Reinforcement Learning
SDE
Stochastic Differential Equation
SPARQL
SPARQL Protocol and RDF Query Language
SPM
Statistical Parametric Mapping
TDD
Test-Driven Development
TF-IDF
Term Frequency–Inverse Document Frequency
TriG
Terse RDF Triple Language with Named Graphs
URI
Uniform Resource Identifier
VAE
Variational Autoencoder
VFE
Variational Free Energy
WCAG
Web Content Accessibility Guidelines
W3C
World Wide Web Consortium
58

## Page 59

10.3
Standard Hypothesis Definitions and Identifiers
The eight hypotheses below define the evaluation rubric used by the LLM-based assertion extractor (extraction
pipeline). Each hypothesis is anchored to its primary domain in the A/B/C taxonomy, but assertions are extracted
from any paper whose abstract relates substantively to the claim. Quantitative results across these hypotheses are
reported in the hypothesis results section.
Table 20: Standard hypothesis definitions tracked throughout the meta-analysis.
The Scope column records the
primary domain in the A/B/C taxonomy; assertions are not restricted to that domain. Wording reflects the prompt
presented to the extraction LLM.
ID
Hypothesis
Scope
H1
FEP Universality: the Free Energy Principle applies
universally to all self-organizing systems, from cells
to ecosystems.
A (Core Theory)
H2
AIF Optimality:
Active Inference agents achieve
principled, near-optimal decision-making under un-
certainty by minimizing expected free energy.
B (Tools)
H3
Markov Blanket Realism:
Markov blankets corre-
spond to real, physically realizable boundaries be-
tween systems and their environments.
A (Core Theory)
H4
Predictive Coding: cortical hierarchies minimize pre-
diction errors via predictive coding, providing a neu-
robiologically realistic substrate for active inference.
C1 (Neuroscience)
H5
Scalability: Active Inference scales to complex, high-
dimensional environments comparable to those ad-
dressed by deep reinforcement learning.
B (Tools)
H6
Clinical Utility: Active Inference produces clinically
useful computational models of psychiatric and neu-
rological conditions.
C4 (Psychiatry)
H7
Morphogenesis:
the FEP explains morphogenetic,
developmental, and self-organizing biological pro-
cesses.
C5 (Biology)
H8
Language AIF: Active Inference provides a viable
framework for language comprehension, production,
and communication.
C3 (Language)
10.4
Glossary of Key Terms
The glossary below defines pipeline-specific concepts, statistical methods, and domain terminology referenced in the
main text. Software package names appear in typewriter font; mathematical objects use the notation defined above.
Where a term has both a colloquial and a technical sense, the technical reading is given.
Table 21: Glossary of key terms used in this manuscript, including pipeline-specific concepts, statistical methods, and
domain terminology.
Term
Definition
Active Inference
A framework in which agents minimize expected free energy to select
actions, unifying perception, learning, and decision-making under the
Free Energy Principle.
Assertion
A directed, confidence-scored claim linking a paper to a hypothesis (sup-
ports, contradicts, or neutral). The basic unit of evidence in the knowl-
edge graph; a machine-extracted classification, not a human verdict.
Continued on next page
59

## Page 60

Term
Definition
Bayesian Mechanics
The formal extension of FEP that grounds Markov-blanket dynamics
in stochastic physics, casting belief updates as gradient flows on a free-
energy potential.
Canonical ID
The unique identifier assigned to each paper during deduplication, fol-
lowing DOI > arXiv ID > Semantic Scholar ID > OpenAlex ID > title
hash.
Checkpoint
A JSON Lines snapshot of LLM extraction progress, recording which
papers have been processed and the resulting assertions, enabling incre-
mental resume after interruption.
Citation-Weighted Score
The hypothesis-level evidence aggregate combining direction, confidence,
and a logarithmic citation weight (Eq. 3).
Compound
Annual
Growth Rate (CAGR)
The constant annual rate that, compounded over the publication window,
takes the cumulative count from the first to the last year (Eq. 10).
Conceptual Nexus Model
(CNM)
The modular knowledge-graph unit used by ResNei; each CNM packages
concepts with provenance and supports longitudinal, latitudinal, and
relational navigation.
Contrastive Divergence
An approximate gradient-based training algorithm for energy-based mod-
els [Hinton, 2002] that truncates the Markov chain used to estimate the
gradient of the log-partition function.
Domain Timeline
Per-domain yearly publication counts visualizing temporal evolution
across the eight tracked categories (A1–A2, B, C1–C5).
Doubling Time (𝑡𝑑)
Years required for cumulative output to double under the prevailing
growth rate (Eq. 9).
Energy-Based
Model
(EBM)
A class of generative models defining 𝑝(𝑥) ∝exp(−𝐸(𝑥)) for an unnor-
malized energy 𝐸. Includes Boltzmann machines, Helmholtz machines,
and VAEs as special or related cases.
Expected
Free
Energy
(EFE)
A scalar combining epistemic (uncertainty-reducing) and pragmatic
(goal-achieving) value, minimized over policies. Decomposes equivalently
into risk + ambiguity or epistemic + instrumental terms [Da Costa et al.,
2020].
FAIR Principles
Findable, Accessible, Interoperable, Reusable: a set of guiding principles
for scientific data infrastructure [Wilkinson et al., 2016]. The pipeline’s
nanopublications satisfy all four.
Free
Energy
Principle
(FEP)
The principle that self-organizing systems minimize variational free
energy—an upper bound on surprise—to maintain their structural in-
tegrity.
Generative Model
A probabilistic model specifying the joint distribution over hidden states
and observations, encoding an agent’s beliefs about how observations are
generated.
Greedy Modularity Maxi-
mization
The
Clauset-Newman-Moore
algorithm
[Clauset
et
al.,
2004]
for
community
detection.
Implemented
via
NetworkX
greedy_modularity_communities;
applied
here
to
the
citation
graph to identify clusters of densely interconnected papers.
HITS
Hub/Authority
Scores
Kleinberg’s mutually reinforcing centrality metrics [Kleinberg, 1999]:
hubs point to many authorities; authorities are pointed to by many hubs.
Helmholtz Machine
A generative model with separate recognition (bottom-up) and genera-
tive (top-down) networks trained by the wake-sleep algorithm [Dayan
et al., 1995]; a direct precursor to the variational autoencoder and the
FEP’s recognition–generation duality.
Incremental Resume
The pipeline’s ability to continue from where a previous run stopped,
loading existing corpus and assertion snapshots and processing only new
papers; controlled by --clear-corpus and --clear-assertions CLI
flags.
Knowledge Graph
A directed graph encoding papers, assertions, hypotheses, and their re-
lationships, serialized in an RDF-compatible format.
Continued on next page
60

## Page 61

Term
Definition
LLM Config
A configuration record specifying the Ollama model name, API URL,
sampling temperature, maximum retries, and retry delay used by the
assertion extractor.
Markov Blanket
A statistical boundary separating internal from external states, defined
as the node set that renders a system conditionally independent of its
environment.
Mean
Year-over-Year
Growth ( ̄𝑔)
Arithmetic mean of (𝑛𝑦−𝑛𝑦−1)/𝑛𝑦−1 across years with non-zero prior-
year counts (Eq. 8).
Named Graph
An RDF graph identified by a URI, enabling multiple graphs to coexist
in a single dataset.
Nanopublications use four named graphs (Head,
Assertion, Provenance, Publication Info).
Nanopublication
A minimal, self-contained unit of publishable knowledge consisting of an
assertion, provenance metadata, and publication context [Groth et al.,
2010, Kuhn et al., 2016].
NMF (Non-negative Ma-
trix Factorization)
A factorization 𝑉≈𝑊𝐻with all factors non-negative, used here for
unsupervised topic discovery (§8.2).
Ollama
A locally hosted LLM server used for assertion extraction; provides repro-
ducibility and avoids external API dependencies [Ollama Team, 2024].
PageRank
A centrality metric originally designed for web-page ranking. In citation
networks, PageRank surfaces influential papers that act as hubs across
otherwise disconnected subgraphs.
Precision
The inverse variance of a probability distribution; in active inference,
precision weighting determines the influence of prediction errors at each
level of a hierarchy.
Predictive Coding
A scheme in which each cortical level passes prediction errors upward
and predictions downward, minimizing local free-energy bounds layer by
layer.
Progressive Parsing
The pipeline’s three-stage JSON recovery strategy for malformed LLM
output: (1) direct parse, (2) strip Markdown code fences and retry,
(3) extract first […] substring. Papers failing all three are logged and
skipped.
Provenance
The recorded lineage of an assertion: source paper, extraction model,
timestamp, and confidence; serialized in the Provenance named graph of
each nanopublication.
Reference Resolution Rate
Fraction of all outgoing references that resolve to another paper inside
the corpus; reported as 7.4% in the present analysis and used as a lower
bound on intra-corpus citation density.
Stochastic
Differential
Equation (SDE)
A differential equation driven by a Wiener (white-noise) process; used
in Bayesian-mechanics derivations of Markov-blanket dynamics.
Surprise
(Self-
Information)
The negative log probability of an observation under the agent’s genera-
tive model; variational free energy is an upper bound on surprise.
Term
Frequency–Inverse
Document Frequency (TF-
IDF)
A weighting that combines normalized term frequency with logarithmic
inverse document frequency (Eq. 7); the standard input to NMF in this
pipeline.
TriG
A serialization format extending Turtle with named-graph support, used
to encode nanopublications as RDF datasets.
Trusty URI
A URI containing a cryptographic hash of its content [Kuhn and Dumon-
tier, 2014], providing verifiable immutability and content-addressable
identification for nanopublications.
Variational
Free
Energy
(VFE)
An upper bound on surprise (negative log evidence) decomposable into
complexity (KL from prior) and accuracy (expected log-likelihood).
Variational Inference
Approximate posterior inference by optimization, replacing intractable
marginalization with optimization of a tractable variational distribution.
Continued on next page
61

## Page 62

Term
Definition
Ward Linkage
A hierarchical clustering method that minimizes total within-cluster vari-
ance at each merge step; used to compute domain-centroid dendrograms
from mean TF-IDF vectors.
Wong Palette
The colorblind-safe 8-color palette of Wong (2011) [Wong, 2011], used
as the standard visualization palette throughout all pipeline-generated
figures.
62

## Page 63

11
References
The bibliography is generated automatically during PDF compilation from references.bib. All citation keys used in
the manuscript (e.g., \citep{friston2010free}) resolve to entries below; unused entries have been pruned. Pandoc’s
--natbib flag injects \usepackage{natbib} and \bibliographystyle{plainnat}, so neither directive appears in
this section or in preamble.md.
63

## Page 64

References
Dmitry Bagaev et al. RxInfer.jl v4.0.0: Real-time and adaptive bayesian inference, 2025. URL https://github.com/R
eactiveBayes/RxInfer.jl.
Vladimir Baulin, Austin Cook, Daniel Friedman, Janna Lumiruusu, Andrew Pashea, Shagor Rahman, and Benedikt
Waldeck. The discovery engine: A framework for AI-driven synthesis and navigation of scientific knowledge land-
scapes. arXiv preprint arXiv:2505.17500, 2025. doi: 10.48550/arXiv.2505.17500.
Rick Bonney, Caren B. Cooper, Janis Dickinson, Steve Kelling, Tina Phillips, Kenneth V. Rosenberg, and Jennifer
Shirk. Citizen science: A developing tool for expanding science knowledge and scientific literacy. BioScience, 59
(11):977–984, 2009. doi: 10.1525/bio.2009.59.11.9.
Jelle Bruineberg, Krzysztof Dolega, Joe Dewhurst, and Manuel Baltieri. The emperor’s new markov blankets. Behav-
ioral and Brain Sciences, 45:e183, 2022. doi: 10.1017/S0140525X21002351.
Théophile Champion, Howard Bowman, and Peter Grünwald. Realizing active inference in variational message passing:
The outcome-blind fixation of belief. Neural Computation, 33(10):2762–2826, 2021. doi: 10.1162/neco_a_01422.
Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science. Behavioral and
Brain Sciences, 36(3):181–204, 2013. doi: 10.1017/S0140525X12000477.
Aaron Clauset, Mark E. J. Newman, and Cristopher Moore. Finding community structure in very large networks.
Physical Review E, 70(6):066111, 2004. doi: 10.1103/PhysRevE.70.066111.
Matteo Colombo and Peggy Seriès.
Free energy: a user’s guide.
Biology & Philosophy, 36(2):1–35, 2021.
doi:
10.1007/s10539-021-09788-0.
Lancelot Da Costa, Thomas Parr, Noor Sajid, Sebastijan Veselic, Victorita Neacsu, and Karl Friston. Active inference
on discrete state-spaces: A synthesis. Journal of Mathematical Psychology, 99:102447, 2020. doi: 10.1016/j.jmp.20
20.102447.
Peter Dayan, Geoffrey E Hinton, Radford M Neal, and Richard S Zemel. The Helmholtz machine. Neural Computation,
7(5):889–904, 1995. doi: 10.1162/neco.1995.7.5.889.
Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey
on RAG meeting LLMs: towards Retrieval-Augmented large language model. arXiv preprint arXiv:2405.06211,
2024.
Zafeirios Fountas, Noor Sajid, Pedro AM Mediano, and Karl Friston. Deep active inference agents using monte-carlo
methods. In Advances in Neural Information Processing Systems, volume 33, pages 11662–11675, 2020.
Daniel Ari Friedman. A template/ approach to reproducible generative research: Architecture and ergonomics from
configuration through publication. Zenodo, 2026a. doi: 10.5281/zenodo.19139090. URL https://doi.org/10.5281/
zenodo.19139090. Active Inference Institute.
Daniel Ari Friedman. docxology/template. Software, GitHub, 2026b. URL https://github.com/docxology/template.
Apache License 2.0. Infrastructure-as-Code system for reproducible computational research.
Karl Friston. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2):127–138, 2010.
doi: 10.1038/nrn2787.
Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology-Paris, 100
(1-3):70–87, 2006. doi: 10.1016/j.jphysparis.2006.10.001.
Karl Friston, Thomas FitzGerald, Francesco Rigoli, Philipp Schwartenbeck, and Giovanni Pezzulo. Active inference:
A process theory. Neural Computation, 29(1):1–49, 2017. doi: 10.1162/NECO_a_00912.
Karl Friston, Lancelot Da Costa, Alexander Tschantz, Conor Heins, and Christopher Buckley. Active inference and
artificial reasoning. arXiv preprint arXiv:2512.21129, 2025. doi: 10.48550/arXiv.2512.21129.
Karl J Friston, Thomas Parr, Yan Yufik, Noor Sajid, Cathy J Price, and Emma Holmes. Generative models, linguistic
communication and active inference. Neuroscience & Biobehavioral Reviews, 118:42–64, 2020. doi: 10.1016/j.neub
iorev.2020.07.005.
Francesco Gregoretti. A C++ implementation of active inference for POMDPs. Neurocomputing, 562:126888, 2023.
doi: 10.1016/j.neucom.2023.126888.
64

## Page 65

Paul Groth, Andrew Gibson, and Jan Velterop. Anatomy of a nanopublication. Information Services & Use, 30(1-2):
51–56, 2010. doi: 10.3233/ISU-2010-0613.
Sarah Hamburg. A guide to DeSci, the latest Web3 movement. Future, 2022. URL https://future.com/what-is-
decentralized-science-aka-desci/. Andreessen Horowitz (a16z) Research.
Conor Heins, Beren Millidge, Lancelot Da Costa, Stephen Mann, Karl Friston, Ozan Catal, Pablo Lanillos, Noor Sajid,
and Alexander Tschantz. pymdp: A python library for active inference in discrete state spaces. Journal of Open
Source Software, 7(73):4098, 2022. doi: 10.21105/joss.04098.
Conor Heins, Beren Millidge, Daphne Demekas, Hrishit Basu, Noor Sajid, and Karl Friston. Collective behavior
from surprise minimization. Proceedings of the National Academy of Sciences, 121(17):e2320239121, 2024. doi:
10.1073/pnas.2320239121.
Conor Heins, Toon Van de Maele, Alexander Tschantz, Hampus Linander, Dimitrije Markovic, Tommaso Salvatori,
Corrado Pezzato, Ozan Catal, Ran Wei, Magnus Koudahl, Marco Perin, Karl Friston, Tim Verbelen, and Christo-
pher Buckley. AXIOM: Learning to play games in minutes with expanding object-centric models. arXiv preprint
arXiv:2505.24784, 2025. Submitted to ICLR 2026.
Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation, 14(8):
1771–1800, 2002. doi: 10.1162/089976602760128018.
Jakob Hohwy. The Predictive Mind. Oxford University Press, 2013. ISBN 978-0-19-968273-7. doi: 10.1093/acprof:
oso/9780199682737.001.0001.
Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. In Proceedings of the 2nd International
Conference on Learning Representations (ICLR), 2014. URL https://arxiv.org/abs/1312.6114.
Rodney Kinney, Chloe Anastasiades, Russell Authur, Isabelle Belber, Jonathan Blaschke, Regan Chiang, Jenna Coffey,
Arman Feldman, Joshua Gruber, et al. The semantic scholar open data platform. arXiv preprint arXiv:2301.10140,
2023.
Jon M Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM (JACM), 46(5):604–632,
1999. doi: 10.1145/324133.324140.
Virginia Bleu Knight, RJ Cordes, and Daniel Friedman. The free energy principle & active inference: a systematic
literature analysis, 2022.
URL https://zenodo.org/records/7449368.
Active Inference Institute. Companion
resources: https://github.com/ActiveInferenceInstitute/Knowledge-Engineering.
Knowledge Pixels. nanopub-js: JavaScript library for nanopublications. Software, GitHub, 2026. URL https://gi
thub.com/Nanopublication/nanopub-js. Version 0.1.0; enables browser-based creation, signing, and querying of
nanopublications.
Franz Kuchling, Karl Friston, Georgi Georgiev, and Michael Levin. Morphogenesis as bayesian inference: A variational
approach to pattern formation and body-plan diversity in biology. Physics of Life Reviews, 33:88–108, 2020. doi:
10.1016/j.plrev.2019.06.001.
Tobias Kuhn and Michel Dumontier. Trusty URIs: Verifiable, immutable, and permanent digital artifacts for linked
data. In The Semantic Web–ISWC 2014, pages 395–410. Springer, 2014. doi: 10.1007/978-3-319-11964-9_25.
Tobias Kuhn, Christine Chichester, Michael Krauthammer, Núria Queralt-Rosinach, Ruben Verborgh, George Gian-
nakopoulos, Axel-Cyrille Ngonga Ngomo, and Michel Dumontier. Decentralized provenance-aware publishing with
nanopublications. PeerJ Computer Science, 2:e78, 2016. doi: 10.7717/peerj-cs.78.
Pablo Lanillos, Cristian Meo, Corrado Pezzato, Ajith Anil Meera, Mohamed Baioumy, Wataru Ohata, Alexander
Tschantz, Beren Millidge, Martijn Wisse, Christopher L Buckley, and Jun Lenz. Active inference in robotics and
artificial agents: Survey and challenges. arXiv preprint arXiv:2112.01871, 2021.
Yann LeCun, Sumit Chopra, Raia Hadsell, Marc’Aurelio Ranzato, and Fu Jie Huang. A tutorial on energy-based
learning. In Predicting Structured Data, pages 191–246. MIT Press, 2006.
Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401
(6755):788–791, 1999. doi: 10.1038/44565.
Michael Levin. Technological approach to mind everywhere: An experimentally-grounded framework for understanding
diverse bodies and minds. Frontiers in Systems Neuroscience, 16:768201, 2022. doi: 10.3389/fnsys.2022.768201.
65

## Page 66

Linhao Li et al. Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge
and Data Engineering, 2024.
Tianyi Liang, Weicheng Zhang, Chao Lin, et al. A survey of scientific information extraction with large language
models. arXiv preprint arXiv:2401.01512, 2024.
Janna Lumiruusu, Daniel Friedman, Shagor Rahman, Vladimir Baulin, and Andrew Pashea. ResNei: Solution design
document. Zenodo, 2025. URL https://zenodo.org/records/16753709. Version 2. Published 2025-08-08.
Beren Millidge. Deep active inference as variational policy gradients. Journal of Mathematical Psychology, 96:102348,
2020. doi: 10.1016/j.jmp.2020.102348.
Beren Millidge. A retrospective on active inference, July 2024. URL https://www.beren.io/2024-07-27-A-Retrospect
ive-on-Active-Inference/. Blog post reviewing the state and trajectory of the Active Inference field.
Beren Millidge, Alexander Tschantz, and Christopher L Buckley. Understanding the origin of information-seeking
exploration in probabilistic objectives for control. arXiv preprint arXiv:2103.06859, 2021.
Beren Millidge, Alexander Tschantz, and Christopher L Buckley. Predictive coding approximates backprop along
arbitrary computation graphs. Neural Computation, 34(6):1329–1368, 2022. doi: 10.1162/neco_a_01497.
Sanjeev V Namjoshi. aif-fep-db: A database of publications related to active inference and the free energy principle.
https://github.com/snamjoshi/aif-fep-db, 2026a. Accessed: 2026-03-20.
Sanjeev V Namjoshi. Fundamentals of Active Inference: Principles, Algorithms, and Applications of the Free Energy
Principle for Engineers. MIT Press, 3 2026b. ISBN 978-0-262-05095-1.
Morteza Nazemi, Sanket Bhatt, and Zafeirios Fountas. Active inference for energy control and planning under partial
observability. arXiv preprint arXiv:2503.07772, 2025. doi: 10.48550/arXiv.2503.07772.
Samuel Nehrer et al. Introducing ActiveInference.jl: A Julia library for simulation and parameter estimation with
active inference models. Entropy, 27(1):62, 2025. doi: 10.3390/e27010062. URL https://www.mdpi.com/1099-
4300/27/1/62.
Ollama Team. Ollama: Run large language models locally. Software, 2024. URL https://ollama.com. Local inference
server for open-weight LLMs.
Thomas Parr, Giovanni Pezzulo, and Karl J Friston. Active Inference: The Free Energy Principle in Mind, Brain,
and Behavior. MIT Press, 2022. ISBN 978-0-262-04535-4.
Marius Pazem, Simon Hangl, and Justus Piater. Free energy projective simulation (FEPS): Active inference with
interpretable policy graphs. In Workshop on Bridging the Gap Between Practice and Theory in Deep Learning,
ICML 2024, 2024. URL https://arxiv.org/abs/2407.05432.
Giovanni Pezzulo, Francesco Rigoli, and Karl Friston.
Active inference, homeostatic regulation and adaptive be-
havioural control. Progress in Neurobiology, 134:17–35, 2015. doi: 10.1016/j.pneurobio.2015.09.001.
Jason Priem, Heather Piwowar, and Richard Orr. Openalex: A fully-open index of scholarly works, authors, venues,
institutions, and concepts. arXiv preprint arXiv:2205.01833, 2022.
Viviana Fernanda Quevedo Tumailli et al. Combining knowledge graphs and large language models: A systematic
literature review. IEEE Access, 2025.
Maxwell JD Ramstead, Paul B Badcock, and Karl J Friston.
Answering Schrödinger’s question: A free-energy
formulation. Physics of Life Reviews, 24:1–16, 2018. doi: 10.1016/j.plrev.2017.09.001.
Rajesh P N Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some
extra-classical receptive-field effects. Nature Neuroscience, 2(1):79–87, 1999. doi: 10.1038/4580.
RDFLib Team. rdflib: A python library for working with rdf. Software, 2023. URL https://rdflib.readthedocs.io/.
Version 7.x.
David H. Rose and Anne Meyer. Universal design for learning. Journal of Special Education Technology, 15(1):67–70,
2000. doi: 10.1177/016264340001500108.
Noor Sajid, Philip J Ball, Thomas Parr, and Karl J Friston. Active inference: demystified and compared. Neural
Computation, 33(3):674–712, 2021. doi: 10.1162/neco_a_01357.
66

## Page 67

Dalton AR Sakthivadivel. On bayesian mechanics: a physics of and by beliefs. Interface Focus, 13(3):20220029, 2023.
doi: 10.1098/rsfs.2022.0029.
Gerard Salton, Anita Wong, and Chung-Shu Yang. A vector space model for automatic indexing. Communications
of the ACM, 18(11):613–620, 1975. doi: 10.1145/361219.361220.
Ryan Smith, Karl J Friston, and Christopher J Whyte. A step-by-step tutorial on active inference and its application
to empirical data. Journal of Mathematical Psychology, 107:102632, 2022. doi: 10.1016/j.jmp.2021.102632.
Theodore D. Sterling. Publication decisions and their possible effects on inferences drawn from tests of significance—or
vice versa. Journal of the American Statistical Association, 54(285):30–34, 1959. doi: 10.1080/01621459.1959.1050
1497.
John Sweller, Paul Ayres, and Slava Kalyuga. Cognitive load theory. Explorations in the Learning Sciences, Instruc-
tional Systems and Performance Technologies, 2011. doi: 10.1007/978-1-4419-8126-4.
Alexander Tschantz, Anil K Seth, and Christopher L Buckley. Learning action-oriented models through active inference.
PLoS Computational Biology, 16(4):e1007805, 2020. doi: 10.1371/journal.pcbi.1007805.
Bo Wen. The missing reward: Active inference in the era of experience. arXiv preprint arXiv:2508.05619, 2025. doi:
10.48550/arXiv.2508.05619.
Christopher J Whyte, Ryan Smith, and Jakob Hohwy.
How do inner screens enable imaginative experience?
A
metacognitive active inference account. Neuroscience of Consciousness, 2025(1):niaf009, 2025. doi: 10.1093/nc/nia
f009.
Mark D. Wilkinson, Michel Dumontier, IJsbrand Jan Aalbersberg, Gabrielle Appleton, Myles Axton, Arie Baak, Niklas
Blomberg, Jan-Willem Boiten, Luiz Bonino da Silva Santos, Philip E. Bourne, et al. The FAIR guiding principles
for scientific data management and stewardship. Scientific Data, 3:160018, 2016. doi: 10.1038/sdata.2016.18.
Bang Wong. Points of view: Color blindness. Nature Methods, 8(6):441, 2011. doi: 10.1038/nmeth.1618.
67


---
*Extraction method: pymupdf*