# Full Text: Abstract book for: 5th Applied Active Inference Symposium

> Extracted from `2025_5thSymposium.pdf`

---

## Page 1

5TH APPLIED
ACTIVE INFERENCE
SYMPOSIUM
ABSTRACT BOOK
NOVEMBER 12-14, 2025
SYMPOSIUM.ACTIVEINFERENCE.INSTITUTE

## Page 2

ACTIVE INFERENCE INSTITUTE
(AII) IS AN OPEN-SCIENCE
INSTITUTE DEDICATED TO
IMPROVING THE ACCESSIBILITY,
RIGOR, AND APPLICABILITY OF
THE ACTIVE INFERENCE
FRAMEWORK.
THERE ARE MANY WAYS TO GET
INVOLVED WITH THE INSTITUTE AND
LEARN ABOUT ACTIVE INFERENCE:
 Active Inference Institute
Application  |   Education   |  Open Science
https://www.activeinference.institute/

## Page 3

5th Applied Active Inference
Symposium
Book Editors: 
Maria Luiza Iennaco & Daniel Friedman
DOI: 10.5281/zenodo.17555267
Part 1
November 12-14, 2025
Part 2
Part 3
Symposium
program

## Page 4

Abstracts
Keynote speakers
Pre-recorded
presentations
Live-streamed pannels
Interactive Workshops
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025

## Page 5

5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
KEYNOTE SPEAKER
Mathematical foundations of
active inference
Karl Friston
University College London
01
Keynote address

## Page 6

The AI Capabilities and
Alignment Consensus Project
(AICACP)
Adam Safron
Allen Discovery Center, Tufts University
The AI Capabilities and Alignment Consensus Project (AICACP) is
a multi-year initiative to reshape the discourse on AI capabilities,
alignment, and regulation. It combines academic publishing with in-
person workshops to bridge divides between different perspectives
on AI. The project also aims to support public engagement through
a podcast series, and potentially an online forum. At the very least,
we hope to provide 2-3 interesting special issues on deep and
timely topics in AI and cognitive science, and to share these
insights with both domain experts and the general public. Most
ambitiously, we hope to start a novel academic society dedicated
to using the highest quality practices for collaborative sensemaking
to figure out what world we are in with respect to increasingly
advanced AI technologies, what this means for science and society
more generally, and what we ought to do about it.
02
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 7

Coherence, Rupture,
Regeneration
Alexander Sabine
University of Portsmouth
The Coherence-Rupture-Regeneration (CRR) offers an early-stage
mathematical framework to explore active inference principles by
incorporating non-Markovian temporal dynamics through
coherence-based memory integration. CRR posits that adaptive
systems maintain identity-through-change via three coupled
operators: coherence accumulation C(x) = ∫ L(x,τ) dτ, discrete
rupture events δ(t-t₀), and regenerative reconstruction Rχ = ∫
φ(x,τ)·e^(C(x)/Ω)·Θ(t-τ) dτ.
A central finding that connects CRR to the Free Energy Principle is
that coherence appears to increase as free energy variational
bound decreases, suggesting that coherence accumulation serves
as a functional proxy for uncertainty reduction. The relationship
emerges naturally from a generalised Euler-Lagrange formulation,
where the standard variational derivative is augmented by
exponentially-weighted memory integrals and impulsive rupture
terms that reset boundary conditions for variational problems.
Unlike standard Markovian models, CRR systems construct their
own temporal structure through coherence fields L(x,ẋ,t), which can
be positive (memory-building) or negative (decoherent). Toy
simulations of CRR show how dynamically transition between
Markovian and non-Markovian regimes based on coherence
density, with rupture events selectively reorganising memory from
the field, rather than simply degrading it. 
03
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 8

This enables metabolised discontinuity; ruptures become
opportunities for adaptive reorganisation, rather than catastrophic
failure.
CRR offers potential as a predictive and diagnostic tool for
understanding systems from neural dynamics to ecological
transitions to cultural evolution, using a single variational calculus
of adaptive temporality. The framework suggests that active
inference agents are fundamentally history-bearing systems whose
temporal structure is actively constructed through coherence-
rupture-regeneration cycles, providing a possible mathematical
language for identity maintenance over time.
04
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 9

Constellatory Cognition: How the
brain can be modelled and use
modelling without having a model
Andrea Hiott
Universität Heidelberg
Recent developments in the study of the hippocampal formation
call old ideas of representation into question and are forcing a
change in the way we understand the study of memory and
navigation, opening the path towards a radical interpretation of the
cognitive map. Through this lens, we can now assess the body’s
cognitive abilities (such as thinking and remembering) through the
same process by which we assess the body’s abilities of
locomotion (such as navigating and wayfinding). In so doing we
benefit from already established ideas in radical embodiment to
move beyond traditional dichotomies of mental and physical. This
gives us an understanding of representations not as ‘findable’ or
‘locatable’ in any hierarchy in the natural world. Rather, the scaling
happens in our assessments: Representations are the ways we
communicate those assessments to one another, and to ourselves;
they are real, and they are interactional.
05
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 10

Making Sense of Uncertainty:
Designing Experiences for
Adaptive Inference
Anna Pereira
Active Inference Institute & Act In Cycle
The theories of active inference and the free energy principle offer
powerful frameworks for understanding perception, learning, and
adaptation, but they remain largely inaccessible outside of
specialized academic communities. This paper introduces a
design-led approach to translating these foundational ideas into
experiential tools that can support adaptive behavior in everyday
life. Drawing on interdisciplinary research, algorithmic metaphors,
and visual communication strategies, we propose a simple, flexible
framework grounded in three interdependent domains: mind, body,
and environment. Within this context, a four-quadrant model is
introduced to engage individuals across varying levels of cognitive
and physical activity, encouraging personalized exploration,
psychological flexibility, and a more coherent engagement with
uncertainty. The approach is supported by analogies to learning
algorithms, including ensemble methods, denoising, and curriculum
learning, which serve as conceptual bridges to guide
understanding. While early-stage and intentionally simplified, these
strategies are intended to evolve iteratively through real-world
application, feedback, and collaborative co-construction. This work
aims to bridge the gap between theory and practice by making the
principles of active inference both comprehensible and actionable
by inviting individuals, researchers, and institutions to explore new
possibilities for adaptation, alignment, and navigating uncertainty.
06
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 11

Approaches to Spatial
Navigation: Embodied and
Cybernetic Underpinnings
Bradly Alicea
Spatial navigation in animals is typically thought of as a learning,
memory, and representational phenomenon based on a process of
consolidating episodic memory. In mammals, spatially explicit
behaviors have been tied to place [1] and grid [2] cells in the
entorhinal cortex, which provide a geometric representation
(abstract map) of allocentric space. In animals more broadly, path
integration provides a more implicit representational mechanism for
tracking an organism’s position [3]. Insects provide an example of
two such mechanisms: using landmarks to calibrate an internal
compass, or by tracking self-motion through cells of the central
complex [4]. Robotic systems use an approach called Generalized
Simultaneous Localization and Mapping (G-SLAM), which utilizes
perception and action to build a representation of the world [5].
Here, we will pull all of these approaches together in terms of their
implications for cybernetics, embodiment, and Active Inference. G-
SLAM provides us with a global representation constructed from a
series of subgraphs [6]. The expansion of spatial representations
occurs through interaction, thus uniting the mechanisms of
entorhinal cell-based cognitive maps and path and positional
integration. Specifically, sensorimotor integration serves to build
and enrich the representation by optimizing and expanding the
graph representing space. 
07
Orthogonal Research and Education Lab; OpenWorm
Foundation; University of Illinois U-C
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 12

08
This type of  representation also leads us to structures consistent with
cybernetic regulatory relations and partially observed Markov models. In a
neural context, these types of  structures allow for the same type of  information
processing on a wide variety of  anatomical substrates.
References:
[1] O'Keefe, John (1978). The Hippocampus as a Cognitive Map. Clarendon
Press.
[2] Moser, E.I., Roudi, Y., Witter, M.P., Kentros, C., Bonhoffer, T., and Moser, M-
B. (2014). Grid cells and cortical representation. Nature Reviews Neuroscience,
15, 466–481.
[3] Biegler, R. (2000). Possible uses of  path integration in animal navigation.
Animal Learning and Behavior, 28, 257–277.
[4] Heinze, S., Narendra, A., and Cheung, A. (2019). Principles of  insect path
integration. Current Biology, 28(17), R1043–R1058.
[5] Al-Tawil, B., Hempel, T., Abdelrahman, A., and Al-Hamadi, A. (2024). A
review of  visual SLAM for robotics: evolution, properties, and future
applications. Frontiers in Robotics and AI, 11, 1347985.
[6] Safron, A., Çatal, Q., and Verbelen, T. (2022). Generalized Simultaneous
Localization and Mapping (G-SLAM) as a unification framework for natural and
artificial intelligences: towards 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 13

Open-source Sustainability:
modeling the social aspects of
production communities
Bradly Alicea
09
In this talk, we will review techniques and approaches to studying
open-source production communities. Production communities are
the ways in which collective behavior is enabled and leveraged in
open-source communities. Of particular interest is the origins of
self-sustaining community production using a Github-like (social
coding and versioned-controlled) platform. Over the past several
years, we have used three different approaches: Partially-
observable Markov Decision Processes (POMDP), Reinforcement
Learning (RL), Agent-based Modeling (ABM), and Large Language
Models (LLMs).To better characterize open-source sustainability,
we must characterize immediate social complexity and long-term
social evolution. Immediate social complexity can be approximated
using POMDPs and RL. The long-term social evolution of social
complexity in projects is modeled in the language of RL and ABMs.
Additional aspects of software production, such as communication
within teams, is approximated using LLM-generated project tasks
and complex network theory. With this collection of tools, we can
think of open-source productivity as a complex system, which may
help us answer questions related to collective behavioral dynamics
and the emergence of community innovation. 
References:
Alicea, B., Ather, H., Chougule, H., McCorkle, B., and Parent, J.
(2023). Open-source Community Sustainability using Agent-based
Models. ResearchGate.
https://www.researchgate.net/publication/369143414_Open-
source_Community_Sustainability_using_Agent-based_Models
Orthogonal Research and Education Lab; OpenWorm
Foundation; University of Illinois U-C
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 14

Combining Hierarchical Active
Inference with Affordance Theory
for Scalable Policy Selection in
Autonomous Drone Navigation
Harshil Shah & Satyaki Maitra
11
This presentation introduces a hierarchical active autonomous
drone navigation project developed within the Microsoft AirSim
simulation environment. The project applies hierarchical active
inference and affordance theory to enable adaptive, efficient, and
interpretable navigation in complex 3D spaces.
It will outline the motivation for building a framework that can
operate under uncertainty, covering key challenges such as
defining action spaces, representing abstract environmental
information, and maintaining computational efficiency for edge
computing. 
Then, there will be a comprehensive breakdown of our architecture.
The presentation will describe how the system fuses LiDAR,
camera, and IMU data to infer a latent suitability state that guides
decision-making by filtering infeasible actions and balancing
exploration and exploitation through Expected Free Energy
minimization. It will also include details on implementation, such as
code snippets and visualizations.
Mission San Jose High School, Active Inference
Institute
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 15

12
Finally, the presentation will cover the strong preliminary results in
our core environment and analyze the strengths and weaknesses
of our architecture. It will conclude by discussing ongoing work
towards more challenging environments, hardware deployment,
and more advanced directions for the architecture.
The website, Over the last few years, many educational and
software resources have been developed for Active Inference. In
this interactive session, open source educational and development
resources will be reviewed and utilized. The focus will be on
presenting (more than) enough entry points, for manual and AI-
augmented learning + coding. Questions, frictions, memes,
dreams, and other input will be curated for continued ongoing
improvement.  has the most up-to-date details on the project and
will continuously be updated.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 16

Modeling the Emergence of
Perspective: An Active Inference
Account of Intentionality
Hongju Pae
13
This presentation applies Active Inference to a foundational
question in consciousness science: how subjective perspective
(subjective intentionality) emerges as a self-organizing process
within living systems. 
Building on the analysis developed in Reflective Analysis on
Empirical Theories in Consciousness (Pae, 2025), I argue that
what may unify multiple mainstream theories of consciousness is
their shared effort to explain the intentional quality of experience -
the fact that consciousness is always from a particular point of
view. 
Within this framework, Active Inference offers a principled way to
model such perspectival structure by treating intentionality as a
latent informational process that modulates qualitative aspects of
experience under continuous minimization of uncertainty about
self-world relations. In doing so, it formalizes the emergence of
subjectivity as an intrinsic feature of autopoietic inference. The
presentation thus applies Active Inference to bridge
phenomenology and computational modeling, illustrating how
intentional stance and perspective formation can arise within
probabilistic generative architectures. 
Research Fellow, Active Inference Institute
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 17

The phenomeno-logical robot
Jean-François Cloutier
Research Fellow, Active Inference Institute
14
I present an update of the Symbolic Cognitive Robotics project I
conduct as an AII Research Fellow.
I am programming a mortal, autonomous robot (a rover with motors
and sensors) so it has felt experiences, makes sense of them, and
evolves its agency to survive in a dissipative world. My project is a
learning-by-doing exploration of phenomenology, active inference
and artificial agency.
The robot will actively evolve a society of mind composed of simple
cognition actors (CAs) organized in an abstraction hierarchy. A CA
observes other, less abstract CAs (its umwelt), updates its beliefs
from patterns it detects, decides which beliefs feel good, and acts
on its umwelt to feel better. A CA uses a logic program it generates
and updates to predict incoming observations and to anticipate the
consequences of its actions.
I am hoping this work will help me answer difficult questions such
as:
Can a robot be made mortal so it can have good vs bad
experiences?
Can a robot's experiences and sense making be entirely its own
and not those of its programmer?
Can there be something it is like to be a robot?
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 18

The missing center: Somatic
belief updating in the global
model.
Jim Freda
Independent Scholar/Practitioner
15
The core node has been undertheorized. The model must itself
contain an internal model. This internal model has an outer
boundary sheet, a Markov blanket of its own. I explore this
somatically in the muscle sense.
In Sherrington’s “neural architecture of the animal as a whole” we
understand that there are three distinct spatial distributions in the
model. We are well acquainted with these three perceptual
systems, except for proprioception. The motoric element of this
deep field is missing from the model. This is a significant barrier to
recognition and action.
Overwhelming surprisal reflects a profound bias in the collective
model, a control bias resulting in a dangerous feedback loop. I
discuss this as exteroceptive bias. There has been a kind of refusal
or repression that we can understand as a double bind in the
blanket politics of the model. Because it contains not just one but
two blankets, necessary for the model to be generative, then we
must theorize their relationship. 
It is the motoric element of the deep field in the model, through
muscle sense, that actively generates the self evidencing needed
for effective belief updating in the face of overwhelming surprisal. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 19

16
This motoric element is the outer surface of the inner self, of the
core node in the model. It both expresses and defends its neural
content and has an important integrative biomechanical role. 
We know this axial system as a series of pits or concavities. These
are major structural junctures, macro-level sensory motor
synapses, each with a single prominent bony servomechanism at
its center. This is a self directed system (of systems) within us and
I describe it’s synergistic kinesiology, allowing us to track its
movements. Please be prepared for guided movement meditations. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 20

The ambiguous status of
attention in predictive models of
cognition
Maria Luiza Iennaco
University of São Paulo / University of Porto
17
Predictive models of cognition, which frame the brain as a
predictive machine, have become prominent in cognitive science.
Despite their promise, one aspect remains under-theorized:
attention. These models attempt to describe (or relate) attention as
the optimization of precision weighting, i.e., the confidence
assigned to sensory signals, but often treat it as a byproduct of
perception and action.
This presentation offers a critical assessment of the potential of
Hierarchical Predictive Coding and Active Inference models to
encompass the different domains of attention – a crucial boundary
phenomenon whose role is central to their entire functioning. To
this end, we will start with a critical survey of the contemporary
status of these models, followed by a more in-depth investigation
of the extent to which they are capable of describing the
particularities of attention.
Following a comparative examination, we discuss the models'
potential and limitations in theoretical and clinical applications.
While they offer appealing philosophical narratives and reasonable
reinterpretations of psychiatric disorders, we believe they currently
lack innovative, testable hypotheses. Finally, we examine important
theoretical restrictions and potential future research areas for
better integrating attention into predictive processing frameworks.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 21

Hyperscanning under Active
Inference for Mental Health
Nicolás Hinrichs
Okinawa Institute of Science and Technology & Max
Planck Institute for Human Cognitive and Brain
Sciences
18
Mental‑health interventions thrive on attunement between therapist
and client, yet quantitative tools for monitoring that attunement are
scarce. 
We present i) a real‑time framework that extends active inference
to dyadic interaction and links it to a geometric hyperscanning
observable suitable for clinical deployment. Each partner in therapy
is modelled as an agent whose self‑model contains a generative
model of the other. Prediction errors in this recursive model
manifest in the topology of their combined neural activity. 
We show ii) how to compute Forman‑Ricci curvature on the
weighted inter‑brain network and track the entropy of its
distribution. 
We introduce iii) a prospective Digital Twin for Psychotherapy
incorporates real-time monitoring of sharp entropy fluctuations,
which timestamp-and ultimately, predict-phase transitions of crucial
in-session behavioral phenomena, such as rupture, co‑regulation
and re‑attunement, to guide the practitioners towards the optimal
outcome: healing. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 22

19
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION
Observer Theory 
Sam A Senchal
Wolfram Institute
This presentation examines a novel category-theoretic extension of
Observer Theory that formalizes how observers construct reality
through constrained sampling of Wolfram's Ruliad—the ∞-groupoid
of all computational processes. The framework introduces
observers as functors that reduce entropy while respecting
boundedness, persistence, and relevance constraints, directly
connecting to Friston's Free Energy Principle.
The proposed hierarchical domain structure—Physical (P),
Valuational (V), Symbolic (S), and Minimally Constrained (M)—with
embedding functors provides a mathematically rigorous approach
to cross-domain causation and information integration. Qualia
emerge as integrated information across domains, offering a
testable formalization of consciousness compatible with Integrated
Information Theory.
Key technical contributions include:
Resolution of discrete-continuous duality through observer-
dependent measure spaces
Formal treatment of cross-domain causation without violating
physical causal closure
Information-theoretic formulation to quantify observer-
accessible novelty
Terminal object (True Infinity) preventing infinite regress in the
observer hierarchy.

## Page 23

20
Panelists will debate whether this synthesis successfully bridges
Active Inference with fundamental physics, the necessity of the
four-domain ontology, and implications for designing artificial
observers. Critical discussion points include the framework's
empirical testability, its relationship to existing theories (IIT, FEP,
Constructor Theory), and whether True Infinity represents
mathematical necessity or philosophical overreach.
This interdisciplinary dialogue promises to advance our
understanding of observers as active constructors of reality, with
profound implications for physics, consciousness studies, and
artificial intelligence.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 24

From Information Economics to
Active Inference: A Unified
Bayesian Mechanics of Choice
Under Uncertainty
Samuel Montañez
Artificial Intelligence PhD student at Universidad
Panamericana (Mexico City)
21
This paper establishes mathematical equivalences between
classical economic decision theory and active inference,
demonstrating that economic behavior emerges from free energy
minimization principles. We show that foundational models in
economics—including Stigler's (1961) optimal search, Simon's
(1955) bounded rationality, Arrow's (1962) value of information,
and Sims' (2003) rational inattention—constitute special cases of
active inference under specific conditions.
Four key theoretical contributions emerge. First, Stigler's optimal
stopping rules correspond to active inference when epistemic
affordance (expected information gain) is excluded, leaving only
pragmatic affordance (expected value). This reveals why empirical
search behavior often exceeds Stigler's predictions: agents
intrinsically value uncertainty reduction beyond its instrumental
benefits. Second, Simon's satisficing emerges naturally when
computational costs are incorporated into the free energy
functional, with aspiration levels representing equilibrium points
where the marginal cost of deliberation equals marginal expected
benefit. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 25

22
Third, Arrow's value of information corresponds precisely to the
Kullback-Leibler divergence between posterior and prior beliefs,
providing the information-theoretic foundation previously absent
from economic theory. Fourth, Sims' rational inattention framework,
which constrains information processing through Shannon channel
capacity, generates the same softmax choice structure that
emerges from variational approximation in active inference—
revealing that economists independently discovered free energy
minimization through information-theoretic reasoning.
This unification has profound implications for behavioral
economics. Documented anomalies—including the Allais paradox,
framing effects, and probability weighting—emerge as optimal
solutions given the brain's hierarchical predictive architecture and
precision-weighting mechanisms. Rather than representing
irrational deviations, these behaviors reflect Bayes-optimal
inference under neural computational constraints. The framework
generates testable predictions linking economic choices to
precision-weighted neural responses in frontal-striatal circuits. By
recognizing that economic agents minimize variational free energy
through coupled pragmatic and epistemic drives, this synthesis
bridges a century of economic theory with contemporary
neuroscience, artificial intelligence, and the physics of complex
self-organizing systems.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 26

Fundamentals of active
inference: A self-guided textbook
for learning and applying active
inference from first principles
Sanjeev Namjoshi
VERSES
23
Over the past decade, research in active inference has expanded
dramatically, generating a wide array of new ideas across diverse
subfields and application domains. Yet, the discipline remains in its
formative stages, despite its potential to transform research and
practice across science and engineering. With the growing
excitement surrounding generative AI and reinforcement learning,
active inference is now uniquely poised to attract attention from
students and researchers in multiple disciplines, setting the stage
for a wave of innovation in both academia and industry reminiscent
of the deep learning surge of the early 2010s. However, despite its
remarkable promise, active inference has yet to achieve the broad
adoption it merits largely due to the steep learning curve and the
lack of accessible, structured introductions in the literature. In this
talk, I present Fundamentals of Active Inference, a textbook
designed to bridge that gap by providing a unified, systematic, and
rigorous foundation for students and researchers entering the field.
The textbook is entirely self-contained and suitable for self-study,
assuming no background outside the basics of probability theory,
linear algebra, and multivariate calculus. Following the release of
the textbook many supplemental educational resources will be
made available including Jupyter notebooks, simulations, video
lectures, and interactive software intended to teach and educate
the concepts of active inference in order to bring these ideas to a
wider audience and induce a paradigm shift leading to the next
phase of both applied and theoretical AI research.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 27

Active Self-Referecing: The concept
of self-reference might be that which
joins complex self-modeling with
reference problems in formal
languages across a spectrum
Sonia de Jager
Erasmus School of Philosophy
24
Referents can be understood as (shared) models encoding
possibilities. All (linguistic, logical, etc.) referents are inherently
ambiguous, and require context for concretization, else they remain
unactualized plans/imaginations/memories. Simple referents with
relatively “easily” concretizable possibility spaces (“cat”) are less
computationally challenging than complex ones (“negation,” or: the
concept of reference itself). We examine the predictive challenge
agents encounter in self-referential “loops”: when reference
“models” reference itself. What predictive function does self-
reference serve? Is it a foundational feature of self-modeling, or a
side-effect hereof? 
The proposal we begin from is that basic self-reference is a
prerequisite for selfhood, where selfhood emerges through the
surveilling and synthesizing of one’s own information-processing.
The loop hereby created is sustained by agents first modelling
themselves as agents to effectively act upon their environment.
Maintaining this recursive process sustains the experience of
continuity through memory and planning (which are not
symmetrical, but intertwined). 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 28

25
We conceptualize this continuity as predictive rehearsal: the
remembering-planning and recursive updating of simulated
experience. For all cognitive systems, lived experience necessarily
involves the continuous rehearsal of self-predictions, but also the
permanent updating hereof. 
We propose that a self-model which is capable of reference (has
linguistic/communicative capacities) updates its predictive self-
referential tension with itself, and this effect is mirrored in the
larger communicative landscape through other processes of self-
reference. What is also set under speculative investigation here is
the concept of self-reference as haecceity: most
linguistic/communicative phenomena are representable as
referents within shared generative models, but reference itself—the
substrate enabling linguistic modeling—cannot be modeled as a
referent within its own system, much like the self remains
(computationally) intransparent to itself, and can therefore be
understood as the ultimate representation of thisness or
individuation. This has implications for understanding
metacognitive prediction limits and the emergence of
perspective/selfhood in active-inferential agents.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 29

Tradescaping for Active
Inference Action Research
Susan Hasty
Independent 
26
TbD
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
PRE-RECORDED PRESENTATION

## Page 30

ALife simulations for multi-agent
alignment of active inference
agents
Adam Safron
Allen Discovery Center, Tufts University
27
We will describe a variety of research questions that we are
preparing to pursue with an integrative simulation environment. We
will implement meta-learning world-modeling agents wherein
intelligence and values co-develop through unified prediction-error
minimization across hierarchical temporal scales. We will deploy
computational architectures based on hierarchical generative
models that perform reciprocal message passing similar to
mammalian nervous systems, enabling flexible construction of self-
world representations with varying temporal depth. Intrinsic drives
for curiosity and empowerment guide exploration, and meta-
learning allows rapid adaptation to new contexts while preserving
established value orientations through precision-weighted attention
to cultural and social cues. Unlike current approaches that treat
capabilities and alignment as separate problems, our approach
embeds value-formation within the fundamental learning process.
We will demonstrate how quasi-Kantian ethics can be made to
emerge naturally from temporally deep policy selection and
generative modeling in ways that require coordinating with others.
This research will involve interdisciplinary collaborations between
individuals with expertise in the psychology of personality (broadly
understood as system-defining attracting states), psychiatry,
multilevel selection and evolutionary game theory, and even basal
cognition (to investigate the generality of these principles).
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 31

5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL
Entropic motivation
Alex Kiefer
Monash Center for Consciousness and Contemplative
Studies; VERSES
28
A key feature distinguishing active inference from more mainstream
AI paradigms like reinforcement learning is the inclusion, in a
principled way/at the ground level, of dimensions of motivation
related to exploration and curiosity. How deep do such “entropic”
principles of motivation go? Are they always in the service of utility
or preference-seeking behavior? I argue that entropy-maximizing
behavior is a sui generis dimension of motivation orthogonal to the
pursuit of reward, and in fact prior to it in the sense that unlike
utility, it is model-agnostic.

## Page 32

Hierarchical Active Inference
Modeling of Social Trust
Dynamics in PTSD: Integrating
Qualitative Phenomenology with
Computational Psychiatry
Andrew Pashea
Active Inference Institute, University of Chicago Harris
School of Public Policy Applied Data Fellow
29
This project explores the dynamics of trust and decision-making in
behavioral task agent-based modeling for individuals with PTSD,
integrating insights from qualitative clinical research, fear
extinction and avoidance behavioral research, and phenomenology
with the Active Inference framework. PTSD is known for involving
difficulties with emotional and physiological regulation in relation to
the aftermath of trauma. This can often manifest behaviorally and
cognitively as, e.g., distrustful or avoidant behaviors, dissociation
or interoceptive disconnect, potential burnout. While previous
research has focused on ethological models in approach-avoid or
explore-exploit tasks, our work emphasizes the human element of
trust in everyday life. We employ the Card Advisor task used in
previous work used to model “delusion” broadly, in the context of
schizophrenia, at the same time aligning our priors more closely to
literature in PTSD and adding a metacognitive layer to the model
for reflective inference about safety. This choice reflects a
translational perspective on psychiatry where the modeling process
can help inform or complement our understanding in cases of
ambiguity when relying upon traditional definitions of
psychopathological disorders.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 33

30
 Our work aims to further interdisciplinary work and perspectives
for aiding clinical practice and understandings of  PTSD. As this
project is still in preliminary stages, with its beginning at the Active
Inference Institute, the primary presenter will include reflections on
the progression of  the project design from a few perspectives,
general modeling flow from a programming perspective, to foster
further conversation and interdisciplinary collaborations.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 34

An invitation to Active Inference
for the Social Sciences
Avel GUÉNIN—CARLUT
University of Sussex / Kairos Research / Active
Inference Institute
31
Active Inference as a model for human cognition has been
productively applied to social cognition for the best part of 10
years, and has more recently been related to open research
questions in anthropology sociology, linguistics, archeology,
political science, and social ontology. In particular, various
researchers have analyzed the consequences of Active Inference
to the following topics: the construction of cultural affordances and
social learning ; the existence and ontological status of collective
and imagined entities ; the nature of social/material organization in
relation to cognition ; and the role of ritual in political power and
social organization.
This literature offers a rich, powerful, and yet elegant approach to
integrate the study of social behavior in the many relevant scales
of organization - from the biological underworld of Levin's "agentive
materials", to the overworld of structural constraints on State
organization. It vindicates and specifies insights from well-
established branches of the social sciences, such as Goffman's
approach to symbolic interactionism, Lakoff's embodied socio-
linguistics, and more generally the pragmatist tradition pioneered
by Dewey, James, Peirce, and Mead (among others) at the turn of
the XXth century. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 35

32
Yet, a problem remains for the Active Inference theorist. This
approach exists at the interface of many different disciplines,
blending of course cognitive and social sciences, but also blurring
the traditional boundaries between sociology and archeology,
design and political theory, anthropology and semiotics, physics
and philosophy. It makes it difficult to pin down to specific fields of
study, and therefore difficult to engage with in the highly siloed
landscape of the academic world. A systematic effort is yet to be
undertaken for the new world to be born, and this livestream
intends precisely to lay out the ground work for new people to join
the effort of building Active Inference for the Social Science. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 36

ActInf in biology: modeling the
environment as an agent
Chris Fields
Tufts University
33
The symmetry of the FEP makes the environment of any active
inference agent an active inference agent. The environment’s
actions on the agent “of interest” are driven by the environment’s
GM, which describes what the environment believes about the
agent. Every action of the environment on the agent can be
considered a communication from the environment formulated by
the environment’s GM. How does this symmetry change the way we
think about and build models in biology, from the molecular scale
to the social or ecological scale? What is the environment “telling”
a polypeptide as it folds into an active protein? What is the
environment telling a differentiating cell during morphogenesis?
What is your environment telling you? What can we infer, as
modelers, about the environment’s GM from the way the
environment acts on the agent of interest?
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 37

From Coordination to Cognition:
Rethinking Agent Communication
Protocols for Active Inference in
the Spatial Web
Denise Holt
AIX Global Media
34
As the global demand for autonomous systems accelerates, a new
class of Agent Communication Protocols (MCP, ACP, A2A, ANP) is
emerging to facilitate interoperability among distributed AI agents.
Yet, these protocols are rooted in architectures that favor reactive
task execution, rather than adaptive, inference-driven behavior.
This presentation explores how these current protocols, though
solving real enterprise needs, are structurally misaligned with the
causal reasoning, real-time adaptation, and generative modeling
required by Active Inference Agents.
I will evaluate each protocol's strengths and limitations in
supporting inference-based decision-making, and contrast their
architectural assumptions with the scale-free, probabilistic
modeling required for Active Inference. More critically, this talk will
introduce how the Spatial Web Protocol (HSTP/HSML) and the
Universal Domain Graph (UDG) offer a dynamic, context-rich
substrate that complements the Free Energy Principle and enables
distributed Active Inference at scale.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 38

35
This session invites the Active Inference research community to
reimagine not just how agents coordinate—but how they learn,
infer, adapt, and act together in an environment built for inference,
not just information exchange. As we shape the future of agent
architectures, it's time to go beyond static message-passing and
toward embodied, semantically grounded communication protocols
that align with the principles of Active Inference.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 39

An Applied Active Inference
Group Practice for
Transdisciplinary Collaboration
Ian Tennant
Anglia Ruskin University
36
Transdisciplinary projects often falter not because of technical
gaps, but because of difficulties in communication, alignment, and
meaning-making across mixed domains. This session introduces a
structured group practice, inspired by therapeutic traditions
including Circling which has recently been developed by Guy
Sengstock and John Vervaeke. Such group work encourages
present-moment awareness, perspective-taking, and reflective
listening. Participants explore the “here-and-now” of interaction
through paraphrasing, impact-checking, and attunement, with the
goal of improving shared understanding and collective insight.
We propose that these group practices can be understood as a
form of applied Active Inference. With…
Generative model alignment - Participants continually update their
beliefs about others’ intentions and perspectives.
Precision calibration (relevance realization) - The practice
highlights which signals deserve confidence and which need
revisiting.
Epistemic action - Asking clarifying questions and mirroring are
ways of actively sampling to reduce uncertainty.
Shared anticipations - Over time, the group develops coordinated
protention and a collective salience landscape.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 40

37
In this session, participants will learn the basics of this group
practice, experience short facilitated practices, and reflect on how
such methods can support transdisciplinary teams in industry. By
framing Circling-like group practice as a live process of Active
Inference, the workshop offers both a theoretical lens and a
practical tool for improving genuine collaboration across
disciplinary boundaries.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 41

Story Graphs: An Exploration of
Use Cases In and Beyond
CogNarr
John Boik
Active Inference Institute, Verses.AI 
38
Story graph is a formalism under development at the Institute’s
CogNarr (Cognitive Narrative) Project that captures and makes
transparent and explicit a storyteller’s belief model regarding some
situation. A story graph might convey a person’s beliefs about what
happened, what might happen, what it means, why it happened,
what future is desired or feared, who was involved, and so on. It is
an intermediate representation between natural language and a
technical implementation (e.g., code in a model) that is readable by
both humans and computers. A story graph conveys information
with less ambiguity than natural language. In the CogNarr setting,
the sharing of story graphs helps to facilitate group cognition at
scale. But story graphs could have applications well beyond
CogNarr, in medicine, industry, media, governance, commerce,
research, education, and civic society, for communication, decision
making, forecasting, model verification and explainability, and
other purposes. This panel will describe the vision behind story
graphs and, with help from the audience, explore possible use
cases that could benefit society.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 42

Adaptive Robotics for the Real
World: A New Paradigm for
Autonomy
Matthew Brown
ThoughtForge 
39
Robotics adoption has long been limited by brittle systems that fail
when the environment changes. A new approach rooted in Active
Inference enables robots to adapt on the fly and learn from their
environment while maintaining stability and precision.
ThoughtForge’s platform integrates these principles into real-time
robotic control, solving key challenges like sim-to-real transfer and
variability in unstructured environments. From adaptive inspection
of large-scale infrastructure to precision manipulation in
manufacturing, this talk highlights real-world applications where
adaptive AI transforms robotics from rigid automation into truly
intelligent systems.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 43

TBD
Michael Garfield
AICACP, Humans On The Loop
40
TBD
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL

## Page 44

5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL
41
An Active Inference Agent for
Asset Pricing
Samuel Montañez
Universidad Panamericana
We propose a unifying, agentic account of financial decision-
making and asset pricing grounded in Active Inference. Rather than
treating uncertainty as exogenous noise to be averaged away, we
model investors as POMDP agents that minimize expected free
energy—balancing extrinsic value (risk-adjusted returns) with
epistemic value (information gain). This perspective reveals
classical approaches—Black–Scholes–Merton (BSM) real options
valuation and stochastic discounted cash flow (DCF)—as limiting
cases that constrain belief dynamics to gradient flows, suppressing
strategic information search. Using a Helmholtz decomposition, we
formalize how Active Inference relaxes this constraint, enabling
curvature-aware belief updates and stable nonequilibrium behavior
on information-geometric manifolds. We illustrate the framework
with a financial agent and a lightweight, model-free baseline,
clarifying when epistemic control improves sequential allocation.
Conceptually, the work links portfolio choice to Bayesian
experimental design and precision (attention) control, offering
testable predictions for neuroeconomics and finance while
providing a reproducible path to engineering belief-driven market
agents.

## Page 45

5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL
42
Money and the Free Energy
Principle
Steph Macurdy
Wolfram Research, Wolfram Blockchain Labs, Quai
Network, UTXO Alliance
An introduction into the topic of cryptocurrency, specifically a two-
currency system called Qi and Quai, and how they relate to energy
and entropy, and how that relates to the Active Inference
community and the free energy principle. (15 to 30 minutes)
Next, a presentation of the math of Active Inference and the Free
Energy Principle. A comparison between two interpretations of the
math, the customary Bayesian interpretations and an alternate
interpretation that describes a dialogue between two systems,
PIx,y) and Q(x). (30 minutes)

## Page 46

5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL
43
Active Neurorobotics: Deep Prior
Preference Learning Guided by
Learnable Goals
Viet Dung Nguyen
Rochester Institute of Technology
Research in active inference (AIF) has led to many improvements
with respect to the robustness of model-based agents that are
applied to both Markov decision processes (MDPs) and partially
observable MDPs (POMDPs). Active inference, essentially, entails
a form of inference and learning that seeks to balance goal-
orienting objectives or prior preferences – exploitative instrumental
signals – with epistemic ones – foraging, exploration-driving
signals. Despite these advances, few studies have investigated the
complexity of prior preferences or leveraged them effectively to
generate informative instrumental signals. Moreover, most AIF
research has overlooked scenarios where goals are provided as
observations, a setting critical for tasks that require explicit goal
querying or instruction-following. In real-world robotic control tasks
– particularly those with complex, dynamically changing goals,
such as language-guided manipulation – standard AIF
architectures are insufficient. In this talk, we will specifically
analyze these limitations by introducing a novel, systematic
framework for deep prior preference model learning guided by
goals or instructions in any modality. Our approach frames AIF
agents in terms of progressively constructing and adapting a prior
preference by leveraging multimodal fusion and queried goals at
each time step, facilitating the dynamic emission of a useful, dense
instrumental signal based on provided instructions for neurorobotic
problem settings. We further study how our prior preference
adaptation scheme makes an AIF agent flexible in challenging
problem contexts, particularly in those that require the querying of
goals related to human instruction or desired states.

## Page 47

5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
LIVE-STREAMED PANEL
44
NGC-Learn V3: A Fast, Modular,
Computational Neuroscience
Library
William Gebhardt
Rochester Institute of Technology
As the field of computational neuroscience expands, the models
being developed and studied are becoming more and more
complex. As a result, the methods utilized to run and simulate
these models take longer and require more resources. This holds
especially true when working with large, stateful systems such as
those that make up spiking neural networks. Furthermore, testing
preliminary ideas and methodological ideas still incur a high cost in
both time and labor to correctly set up all of the moving parts,
generally resulting in highly specialized, brittle code that will need
to be stripped down in order to be reused for another, later idea. In
the Neural Adaptive Computing (NAC) Lab, we have designed
NGC-Learn to be a library that handles key experimental concerns
related to optimization and reusability. The library allows users to
focus on what really matters -- the research and development of
novel methods and models -- without worry as to the simulation
and design overhead. Join us in this high-level tech demo of the
latest version of NGC-Learn, covering its design patterns and
modularity as well as the internal processes that it uses to speed
up computation.

## Page 48

Using Active Inference for the
Management of Distributed
Energy Resources in CityLearn
Kobus Esterhuysen
LearnableLoopAI.com
45
Active inference offers a powerful framework for managing
distributed energy resources (DERs) because it integrates
perception, learning, and action in uncertain environments, using
probabilistic models to make adaptive decisions about resource
allocation and control. This approach enables energy systems to
self-optimize and dynamically respond to changing demand,
generation fluctuations, and grid conditions, improving energy
efficiency, reliability, and long-term resilience. By leveraging the
capabilities of active inference, operators can better coordinate
diverse DER assets without full system knowledge, efficiently
balancing supply and demand while supporting the integration of
renewables and distributed generation.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
INTERACTIVE WORKSHOP

## Page 49

Interactive Math Tools for Active
Inference Education
Octopus
Eight Arms Nine Brains 
46
To help students reach their goals, the Teacher must be open to
and welcome feedback. This session will discuss tools and
strategies for eliciting feedback in various forms. 
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
INTERACTIVE WORKSHOP

## Page 50

47
In this session, we will take a guided walk through the emerging
landscape of the 3D internet — the metaverse — and explore how
principles of Active Inference and the Free Energy Principle can
inspire the design of meaningful, interactive experiences. At
Numen.games, we have spent the past several years
experimenting with how these theoretical frameworks can be
applied in practice, consciously embedding them into gamified,
educational environments.
Our partnership with the Active Inference Institute allows us to
share these spaces with a broader community, creating
opportunities for exploration, learning, and collaboration.
Participants will not only hear about our journey but will also
engage directly with interactive demonstrations inside a virtual
environment. By moving beyond theory into lived, embodied
experience, we aim to show how Active Inference can inform the
way we build and inhabit digital spaces.
The session will be highly participatory: attendees will be invited to
join us inside a 3D online environment, interact with one another,
and reflect together on how the principles of Active Inference
shape perception, action, and communication in real time. This
interactive component ensures that the session is not just about
describing ideas, but about experiencing them.
Ultimately, we invite participants to imagine with us how the
metaverse can become more than entertainment: a space for
education, scientific exploration, and collective sense-making —
grounded in the principles of Active Inference.
Active Inference & RPGs
PabloFM
Role Playing Games
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
INTERACTIVE WORKSHOP

## Page 51

Workshop on ActiveInference.jl.
Peter Thestrup Waade
Aarhus University 
48
In this presentation, we will describe recent updates to
ActiveInference.jl, as well as how it connects to other software
packages in the larger active inference and Bayesian inference
ecosystem, such as RxInfer.jl and Turing.jl. ActiveInference.jl has
been extended to be able to accommodate other generative models
than the discrete POMDP classically used in the field, such as
generalized filtering, hierarchical gaussian filtering, differential
equation models or predictive coding networks. The package also
allows for using different types of perceptual inference methods
(such as amortized inference or different variants of variational
Bayes), and different variants of the actions selection process
(sophisticated inference etc.). We have aimed at making the
package easy to use - and easy to extend and develop for
contributors, in the hope that it can help facilitate iterative
development in the community. Finally, owing to Julia’s native
autodifferentiability, ActiveInference.jl is invertible, so that it can be
used for computational phenotyping (or cognitive modelling, as the
method is more broadly known), such that inference can be drawn
about unknown generative models held by systems of interest,
such as human participants in psychiatric experiments.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
INTERACTIVE WORKSHOP

## Page 52

Consciousness and Active
Inference
Robert Worden
Active Inference Institute
49
     Recent work on consciousness implies that a pure computation
cannot be conscious. This is because all information in a computer
is encoded, whereas consciousness is not encoded. This means
that the neural computations of Active Inference are not conscious.
Consciousness could arise from an analogue model of 3-D space in
the brain, held as a wave in the thalamus. If it does, how do the
neural computations of Active Inference couple to the wave?
Possible couplings between Active Inference and consciousness
are described. The workshop will discuss research projects to
investigate them.
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
INTERACTIVE WORKSHOP

## Page 53

Thermodynamics
Patrick Huembeli & Maxwell Ramstead
49
5TH APPLIED ACTIVE INFERENCE
SYMPOSIUM 2025
INTERACTIVE WORKSHOP

## Page 54

*[Page 54 appears to be blank or image-only]*

## Page 55

5TH APPLIED
ACTIVE INFERENCE
SYMPOSIUM
ABSTRACT BOOK


---
*Extraction method: pymupdf*