# Extracted text

Source: Zenodo record23986505; PDFSHA256 cbb841093377dbeacdae95e8ac035c12be81d41bd59c66bdb6edf679febbf4a8.
Method: PyMuPDF 1.28.2; all62pages in original content order.

## Page 1

2026 Active Inference Symposium —
Presenter Program
Working Program
November 12–13, 2026 (UTC)
https://symposium.activeinference.institute/
blanket@activeinference.institute
DOI: 10.5281/zenodo.23986505
Active Inference Institute
Contributing presenters
Bradly Alicea
Dan Corva, Kevin Backhouse
Hongju Pae
Karl Friston
Alison Gopnik
Bianca Avanzo
Ed Ober
John Boik, Glen EP Ropella
Luca Possati
Majid D. Beni
Mike Levin
Montijn van Loo
Ryan Smith
Sergio Rubin
Takuya Isomura
Ted J. Vrioni
Ahmad Nazzal
Alexander Sabine
Andrew Pashea
Anna Pereira
Avel Guénin—Carlut
Bo Lee
Bruce Tisler
Cleber Gomes
Colin Turfus
Daniel Dragolich
Daniel Friedman
Diaconescu Adrian-Stelian
Harshil Shah, Andrew Pashea
Harshil Shah, Satyaki Maitra
Holly Grimm, John H. Clippinger,
Daniel Ari Friedman
Hyojun Choi
Jean-François Cloutier
Lancelot Da Costa
Matthew Brown
Michael Lennon
Michael Walker
Nick Hay, Adam Safron
Patricia Craja
Paulo Santiago
Rajkaushik Borgohain
Robert Worden
Sheena Yap Chan
Stell
Steph Macurdy
Tao Gong
Tiga Kailimai
Tran Hoang Linh
Viraj Deshwal
Vitalii Shapovalov
Vitória de Souza
Xin Sun
Book of Abstracts

## Page 2


## Page 3

Contents
About this program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
1
An in-depth discussion on “World models in natural and artificial intelligence”
. . . . .
6
2
Build a curious rocket: a hands-on continuous-state AIF session with cpomdp . . . . . .
7
3
A Modular Perspective Layer for Active Inference Agents . . . . . . . . . . . . . . . . .
8
4
Active Inference and Free Energy Principle in 2026 . . . . . . . . . . . . . . . . . . . .
9
5
Active Inference, Bayesian Causal Learning and Empowerment in RL. . . . . . . . . . . 10
6
Root Frequency Theory: An Integrative Framework for the Continuity of Lived Experience 11
7
Working Groups Sessions (2 sessions: Opening & Charter; Outcomes & Next Steps)
. . 12
8
CogNarr Story Graph: A Formal Semantics for Narrative as a Belief Model
. . . . . . . 13
9
Active Inference and the Evolution of Mind and Society . . . . . . . . . . . . . . . . . . 14
10
From Self-Evidencing to Global Governance: Active Inference and the Dynamics of Justice 15
11
“Evolution, Emergence, Prediction.” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
12
A New Politics: From Linear to Sorted Politics Through the Policy Map
. . . . . . . . . 17
13
Explicating and Comparing Prediction Architectures of Theories of Consciousness and
its Application to INTREPID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
14
Recent advances in empirical applications of active inference in computational psychiatry 19
15
An elementary form of agency at planetary-scale . . . . . . . . . . . . . . . . . . . . . . 20
16
Reverse engineering generative models from zebrafish decision-making . . . . . . . . . 21
17
Knowthyself — A code-free active inference workbench . . . . . . . . . . . . . . . . . . 22
18
A Lexical T‑Maze for Active Inference: A Proof‑of‑Concept Model of Context Seeking
Under Ambiguity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
19
FEP and Active Inference through the lens of Winnicott’s Potential Space . . . . . . . . 24
20
Active Inference Tutorial: Perspectives, Correlates, and Future Directions for Graphical
Brains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
21
Inferring the Unseen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
22
A precis of two principles - Towards a relational perspective on the FEP . . . . . . . . . 27
23
Maladaptive Daydreaming as a Precision Trap: An Active Inference Framework, Its
Testable Core, and Evidence from Daily Diaries
. . . . . . . . . . . . . . . . . . . . . . 28
24
The Benefits of Circuit Convolution: behavioral complexity arises from developmentally-
simulated nonlinearity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
25
The Minimal Necessary Basis for Primitive State-of-Affairs Inquiry . . . . . . . . . . . . 30
26
Artificial Sentience Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
27
Realms of cognitive engagement in Active Inference . . . . . . . . . . . . . . . . . . . . 32
28
NUMEN: Bare metal inference
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
29
GNN in 2026: Generalized Notation Notation as a Specification Layer for Applied Active
Inference
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
30
Stimming as a Temporary Homeostatic Buffer and Modulator in Autism Spectrum Dis-
orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
31
Partner-Specific Affective Precision in Social Active Inference
. . . . . . . . . . . . . . 36
32
VAMAI: Variational Affordance Mapping and Active Inference for Training-Free UAV
Navigation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
33
Coupling energy-system models with a seam you can read and an agent you can audit:
Active Inference in the NEST programme . . . . . . . . . . . . . . . . . . . . . . . . . . 38
34
Adaptive Prosociality Through Allostatic Regulation in Social Dynamics: A Simulation
Study
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
35
Who knows
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
36
Toward Active World Modeling
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
37
Active inference and design
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
38
Unlocking Robotic Dexterity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
39
Today’s Dialogue, Tomorrow’s Inheritance: Active Inference for Networks That Learn
Together.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
40
Active Inference, Information, Human Evolution . . . . . . . . . . . . . . . . . . . . . . 45
41
Solving the Relevance Problem: A General Relevance Criterion in Active Inference . . . 46
42
The Mark of the Moral and the Shared Author Hypothesis: An Active Inference Approach
to Moral Objectivity
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
43
Commongrounds AI: alife-inspired virtual environments for embodied active inference
agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
44
Percepia: Perceptual Phenotyping from Conversation
. . . . . . . . . . . . . . . . . . . 49
45
Congruent Interaction Geometry
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3

## Page 4

46
The Screen as a Markov Blanket: An Active Inference Account of Feed-Driven Narrow-
ing, and a Personal Instrument to Measure It . . . . . . . . . . . . . . . . . . . . . . . . 51
47
Language, Self and the Shadow Audience
. . . . . . . . . . . . . . . . . . . . . . . . . 52
48
When Agency Stalls: The Human Readiness Gap . . . . . . . . . . . . . . . . . . . . . . 53
49
Where the Theories Stop: What Free Energy and Integrated Information Actually Buy
You in a Running System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
50
Entropy Minimization is Maximal Order
. . . . . . . . . . . . . . . . . . . . . . . . . . 55
51
What bounds the surprisal? Zones of Bounded Surprisal, self-maintenance and the in-
formational turn in anthropogenesis
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
52
What becomes available next? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
53
Dynamic Action-Space Adaptation via Contrastive Transition Adaptation in Active Infer-
ence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
54
A derived self . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
55
ROIF and Active Inference: The Formal Boundary of History-Conditioned Anticipatory
Systems.
History-dependent reconfiguration, structured memory, and predictive pre-
configuration in prestressed complex systems
. . . . . . . . . . . . . . . . . . . . . . . 60
56
Modelling Language Comprehension with Embodied Active Inference . . . . . . . . . . 61
57
From Lived Experience to Mathematical Mapping: A Computational Phenomenology of
Trauma Recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4

## Page 5

About this program
Event dates: 2026-11-12 to 2026-11-13.
This working program collection contains approved public abstracts and explicitly marks entries
awaiting an abstract. Inclusion does not assert that a presentation has occurred or that a recording
has been received.
Presenter details reflect public self-reports; unavailable fields are omitted.
5

## Page 6

1
An in-depth discussion on “World models in natural and
artificial intelligence”
Presenter: Bradly Alicea
Contribution: Interactive Workshop
Affiliations: Active Inference Institute (SAB)
Abstract
This session will be a summary and discussion group for the special issue ”World Models in Natural
and Artificial Intelligence”, published in the Royal Society A journal in May 2026. We will offer a
summary of the special issue content, along with a discussion of Active Inference related topics
and future directions from the work presented in the issue.
6

## Page 7

2
Build a curious rocket: a hands-on continuous-state AIF
session with cpomdp
Presenter: Dan Corva, Kevin Backhouse
Contribution: Interactive Workshop
Abstract
In the 1960s, Rudolf Kalman wrote down a filter that would go on to navigate Apollo to the moon.
It has been the workhorse of continuous-state estimation ever since, and if you have ever built a
continuous active inference agent in the linear-Gaussian regime, you have built one too, whether
you intended to or not. Bolt a goal onto it and you get a very good controller that is not curious in
the slightest. Koudahl, Kouw and de Vries proved why in 2021: under a fixed sensor the epistemic
value is identical for every action, so expected free energy quietly reduces to Linear Quadratic
Regulator (LQR). In this workshop, in about ten lines of Python that will look familiar to anyone
who has used pymdp, we are going to build a rocket! Then, we add the interesting part, curiosity,
by making a surgical, biologically realistic change: adding sensor noise that varies based on the
rocket's position. We will see how the flight path bends when we add artifacts like a navigation
beacon or a blackout zone, and explore the science as to why. You will write every line yourself
in a Colab notebook and do not worry there will be plenty of scaffolding from the cpomdp team.
Code hints will drop in on a timer, and a reference cell for anyone who gets stuck or would rather
just see the finished steps. We will build this in three steps: the 1960 rocket, a nav beacon it can
fix its position by, and a blackout zone. Everything we use runs on the PyPI release of cpomdp and
anything you build in the notebook is yours to keep and continue working on. Prerequisites: com-
fortable editing a few lines of Python. Familiarity with pymdp helps but is not required. Nothing
to install; a browser and a google account is enough. If you would like to use your own IDE feel
free but we may not be able to provide 1-on-1 support if you get stuck.
7

## Page 8

3
A Modular Perspective Layer for Active Inference Agents
Presenter: Hongju Pae
Contribution: Interactive Workshop
Abstract
Most agent architectures collapse two things that phenomenology keeps apart: what the agent
should do immediately, and what kind of world the agent takes itself to be in. This session intro-
duces “perspective” as a distinct architectural layer that carries the second. Perspective is a slowly
evolving, history-sensitive latent state that gates how observations are encoded. Importantly, it
is decoupled from policy gradients and is never itself an optimization target. Because it is decou-
pled, it is portable and can be attached on top of an existing policy-optimizing agent rather than
requiring a bespoke design. The first part of the session develops the conceptual frame: how a
phenomenologically motivated condition translates into a concrete architectural commitment, and
why that commitment is orthogonal to belief-state inference and policy selection. I then trace the
capabilities that this single layer has been shown to support across a sequence of studies, including
directional hysteresis, history-dependent perceptual reorganization, and body-grounded conative
orientation, treating these as unified aspects of a single portable module. There will then be a live
demonstration attaching the perspective layer to an active inference agent, such as a hierarchical
Bayesian / discrete state-space model, and running it in parallel with its unaugmented counterpart
in the same environment, so participants can see exactly what the layer adds and where it plugs
in. The session closes with a structured discussion in which participants are invited to identify
candidate insertion points for a perspective layer in their own active inference agents, whether
built in PyMDP, RxInfer, or custom frameworks. The aim is to leave participants able to reason
about whether and how a subjective perspective could be added.
8

## Page 9

4
Active Inference and Free Energy Principle in 2026
Presenter: Karl Friston
Contribution: Keynote & Panel
Abstract
No program abstract has been published for this presentation yet.
9

## Page 10

5
Active Inference, Bayesian Causal Learning and Empower-
ment in RL.
Presenter: Alison Gopnik
Contribution: Livestreamed panel or group session
Abstract
Cognitive scientists have applied advances in our formal understanding of causation in philosophy
and computer science, particularly within the Causal Bayes Net formalism, to understand human
causal learning. This formalism is also closely related to ideas in active inference. In parallel,
reinforcement learning researchers have developed the idea of an intrinsic reward signal called
“empowerment”. An agent is rewarded for maximizing the mutual information between its actions
and their outcomes, regardless of the external reward value of those outcomes. “Empowerment”
may be an important bridge between classical Bayesian causal learning and reinforcement learning
and may help to characterize causal learning in humans and enable it in machines
10

## Page 11

6
Root Frequency Theory: An Integrative Framework for the
Continuity of Lived Experience
Presenter: Bianca Avanzo
Contribution: Livestreamed panel or group session
Abstract
How do we keep a coherent sense of self across time — and can that process be measured in
real time, on consumer hardware, before it breaks down? Root Frequency Theory (RFT) proposes
that experiential continuity is not a default state but an ongoing biological achievement requir-
ing coordination across five nested regulatory layers: environmental constraints (C0), autopoietic
regulation (C1), predictive neural integration (C2), narrative self-organization (C3), and the conti-
nuity of lived experience (C4). Fragmentation is defined as a self-reinforcing cascade of cross-scale
misalignment. The cascade is bidirectional: structured self-knowledge introduced at C3 — opera-
tionalized as epistemic alignment (Ke) — is hypothesized to initiate upward regulatory stabilization.
To test this, I developed the M-RFT metric: M = (Φ · Ke) / F, where Φ indexes autonomic flexibility
via HRV, Ke composites interoceptive accuracy and narrative coherence, and F captures regula-
tory burden. I then built Root Extension — an iOS application using iPhone and Apple Watch — to
administer this protocol ecologically and deliver personalized epistemic scaffolding upon session
completion. The N=1 longitudinal study ran 321 sessions across 23 active days (April 16–May
14, 2026), with the researcher as the sole subject. Key findings: mean M = 3.79 (Mdn = 2.14),
intra-session regulatory recovery — FRAGMENTED to RESILIENT within the same calendar day
— observed on 34.8% of active days. This presentation introduces the theoretical architecture, the
metric, the app protocol, and the preliminary data — and invites the active inference community
into the next phase: a formal N=20–40 validation trial planned for 2027.
11

## Page 12

7
Working Groups Sessions (2 sessions: Opening & Charter;
Outcomes & Next Steps)
Presenter: Ed Ober
Contribution: Livestreamed panel or group session
Abstract
AII proposes to facilitate / host active inference working groups as explained in our paper “Coordi-
nating the Field”, with a field group and one or more domain groups. The first symposium session
(Opening & Charter, November 12) will introduce the concept, recap the paper and reason for the
groups, and solicit suggested domains and participants. It will also review the proposed charter
and consider amendments, with versioning per group, considering nominations for chair(s) and
secretary(ies), with a goal of adopting the charter(s) during the symposium if possible and estab-
lishing post-symposium meeting cadence and work plan. The second session (Outcomes & Next
Steps, November 13) will report back to the main session, recap what was accomplished during
the symposium and announce any working groups formed and their first year goals established.
12

## Page 13

8
CogNarr Story Graph: A Formal Semantics for Narrative as
a Belief Model
Presenter: John Boik, Glen EP Ropella
Contribution: Livestreamed panel or group session
Abstract
People naturally express their beliefs about a topic, situation, or problem in the form of narra-
tive. CogNarr is a soon-to-be-released, open-source, interactive tool that lets a writer transform
narrative into an explicit, formal semantic graph, which they can inspect and correct until it faith-
fully reflects their belief model. A CogNarr story graph captures a narrative's argument structure,
its timing, and the modality of its claims, including predictions and conditionals, together with
the uncertainty attached to each. In general it expresses the verifiable meaning of what a per-
son wishes to say, while reducing the ambiguity of natural language. As such, it could be of use
in modeling the beliefs that active-inference processes operate over, and as a framework within
which active-inference methods could be employed to infer semantic meaning from text. Because
it makes a person's beliefs explicit, transparent, and shareable, CogNarr has potential wherever
consequential beliefs are expressed as narrative, from medicine and governance to public polling
and accident reporting. As well, CogNarr story graphs support accurate agent-agent communica-
tion. The collective-cognition aim behind the project is alignment of shared belief models during
decision-making processes in the large-group setting. The talk will be a preview of features and
capabilities, with some discussion of potential applications.
13

## Page 14

9
Active Inference and the Evolution of Mind and Society
Presenter: Luca Possati
Contribution: Livestreamed panel or group session
Abstract
This panel explores how active inference and the free-energy principle can contribute to a unified
account of psychological and social evolution. It examines how cognition, affect, social interaction,
cultural practices, and institutions emerge through processes of prediction, adaptation, learning,
and uncertainty reduction. By bringing together perspectives from cognitive science, philosophy,
anthropology, and social theory, the panel asks how individual and collective forms of agency co-
evolve and how shared models of the world shape both minds and societies.
14

## Page 15

10
From Self-Evidencing to Global Governance: Active Infer-
ence and the Dynamics of Justice
Presenter: Majid D. Beni
Contribution: Livestreamed panel or group session
Abstract
Active inference frames biological and cognitive agency as the imperatives of self-evidencing—
maintaining systemic integrity by minimising variational free energy within an open, stochas-
tic environment. Under this formalism, agents, from single cells to complex organisms, persist
by regulating their internal states against entropic dissipation, enacting a Markov blanket that
partitions self from world. This imperative entails sophisticated inference, balancing epistemic
drives against pragmatic preferences, thereby constituting adaptive agency.
This talk extends
this foundational picture of selfhood to the domain of political communities and global justice the-
ory, demonstrating that the same dynamical constraints that govern individual persistence scale
hierarchically to structure collective organisation and inter-state conduct. Drawing on the Free
Energy Principle, political communities are modelled as open, self-organising systems embedded
within a nested hierarchy of interacting Markov blankets. Their stability is not intrinsic but re-
lational, contingent upon continuous exchange of information, resources, and institutional con-
straints with their environment.
Within this multiscale architecture, global justice emerges as
an active regulatory imperative: a homeostatic constraint that maintains a nonequilibrium steady
state across coupled societies. We formalise this through the expected free energy decomposition,
where just policies balance risk—adherence to shared normative constraints such as human rights
and non-aggression—against ambiguity—the active resolution of uncertainty arising from struc-
tural interdependence. The normative implications are acute. We identify a pervasive pathology
of self-organisation, wherein dominant or insular political regimes mis-specify themselves as closed
systems. By suppressing external coupling and attenuating sensory evidence, they optimise local
coherence at the expense of systemic stability. Drawing parallels to oncogenic cellular prolifera-
tion and psychotic cognitive inference, we argue that such closure systematically underestimates
global risk and propagates destabilising volatility.
Ultimately, we contend that justice is not a
static distributive outcome nor a single-level contractual procedure, but the dynamic property of a
multiscale system that successfully synchronises its trajectories under persistent uncertainty. This
offers a unified, first-principles framework for linking the philosophy of agency to the normative
demands of a deeply interconnected global order.
15

## Page 16

11
“Evolution, Emergence, Prediction.”
Presenter: Mike Levin
Contribution: Livestreamed panel or group session
Abstract
No program abstract has been published for this presentation yet.
16

## Page 17

12
A New Politics: From Linear to Sorted Politics Through
the Policy Map
Presenter: Montijn van Loo
Contribution: Livestreamed panel or group session
Affiliations: Tilburg University
ORCID: 0009-0003-0249-3240
Presenter background
I am an econometrist gone rogue. Instead of throwing mathematics at finance, I throw Active
Inference mathematics at philosophy. See for the consciousness, morality, politics, gender, free
will and God applications, with the poem uniting them all.
Abstract
Political disagreement often reflects competing worldviews: different beliefs about how the world
works and what will happen under alternative policies.
Yet political debate proceeds linearly,
through proposals, responses, and counterresponses.
The underlying reasoning and evidence
remain fragmented, assumptions can escape scrutiny, and criticism loses its target as debate
moves on.
Consequently, incompatible worldviews can persist without their disagreements be-
ing resolved. This talk proposes a new politics, motivated by Active Inference, which provides
the mathematical object around which political debate can be organized: the policy map. This
map expresses a worldview through predicted outcomes under alternative policies.
Making it
the persistent object of debate allows criticism to retain a stable address rather than disappear
in the rhetoric of successive exchanges. The mathematics of the policy map distinguishes three
types of criticism: belief criticism, challenging beliefs about current states or the consequences of
policies; outcome incompleteness, identifying omitted consequences; and policy incompleteness,
identifying missing alternatives. These provide the horizontal sorting of political debate. Vertical
sorting organizes the reasoning beneath each prediction, allowing criticism to reach its underlying
assumptions and evidence. Unresolved objections remain visible until answered or incorporated
through revision, potentially changing the predicted outcomes and preferred policies. In this way,
Active Inference offers a normative application that goes beyond describing politics as if political
actors performed Bayesian inference. It motivates transforming the very structure of political de-
bate, replacing rhetorical exchanges with cumulative criticism and incentivizing a shared search
for the worldview that best fits the available evidence. The ultimate aim is a politics in which
criticism can actually bite, transforming not only political debates, but the world shaped by its
decisions. For the only path forward is critique without exception.
17

## Page 18

13
Explicating and Comparing Prediction Architectures of
Theories of Consciousness and its Application to IN-
TREPID
Presenter: Montijn van Loo
Contribution: Livestreamed panel or group session
Affiliations: Tilburg University
ORCID: 0009-0003-0249-3240
Presenter background
I am an econometrist gone rogue. Instead of throwing mathematics at finance, I throw Active
Inference mathematics at philosophy. See for the consciousness, morality, politics, gender, free
will and God applications, with the poem uniting them all.
Abstract
Consciousness: The Channel Framework and its Application to INTREPID The Channel Framework
treats theories of consciousness as relating physical and experiential events, with the minimal re-
lation being predictive. The Channel Framework embeds this psychophysical relation within the
full experimental prediction architecture, connecting physical processes, conscious content, intro-
spective access, and observable outcomes. Expressed through probabilistic channels, these archi-
tectures can be interpreted as generative models of experimental predictions, providing a common
language in which different theories can be explicated and compared. I apply the framework to
INTREPID’s adversarial comparison of Integrated Information Theory and predictive-processing
theories of consciousness, including Neurorepresentationalism and the Active Inference theory
of consciousness (AI-C). The framework asks whether their prediction architectures can be faith-
fully aligned, which theoretical differences actually generate their contrasting experimental pre-
dictions, and whether these differences carry the particular claims the experiments aim to test.
Building on this analysis, the Channel Standard develops criteria for attributing evidence from
competing theory predictions to specific theoretical claims. The broader aim is to move beyond
theory-level comparison toward methodological concensus.
18

## Page 19

14
Recent advances in empirical applications of active infer-
ence in computational psychiatry
Presenter: Ryan Smith
Contribution: Livestreamed panel or group session
Abstract
In this talk, I will review recent studies and results using active inference and related modeling
approaches to study mechanisms of normative cognitive-affective processes and associated neu-
ral responses, and how these appear to be affected in different psychiatric conditions. I will also
discuss contexts in which active inference appears to offer potentially unique resources for mod-
eling these processes. Conversely, I will discuss contexts in which active inference may overlap
with simpler modeling approaches and related issues centered around the level of mechanistic
description one is interested in studying.
19

## Page 20

15
An elementary form of agency at planetary-scale
Presenter: Sergio Rubin
Contribution: Livestreamed panel or group session
Abstract
The claim that climate homeorhesis—non-equilibrium steady-state dynamics that underlie Earth’s
long-term habitability—depends on an integrated biosphere–climate system capable of acting on
its own is central to debates surrounding the Gaia hypothesis.
This paper presents a proof of
concept suggesting a minimal form of planetary agency based on a simple biosphere–climate sys-
tem model operating through active inference. This means that the internal states of the system
(biosphere) will appear to infer shifts in net incoming solar radiation (external states) through
its boundary states, i.e. by integrating past and current surface temperatures (sensory states)
and actively changing greenhouse forcing (active states) to maintain climate trajectories within
geophysiological bounds. When surface temperature is altered by solar fluctuation changes, the
system adjusts greenhouse and albedo parameters to restore expected climate trajectories, anal-
ogous to correcting sensory prediction errors in biological systems. Perturbing active states (e.g.,
imposed changes in albedo) does not reduce inference precision. Inference robustness decreases
when the system’s degrees of freedom increase. Active inference falls, and thus climate home-
orhesis fails when there is no biosphere activity (lifeless planet) or when it is decoupled from the
climate system. Predictive behaviour operates within the range of solar constants consistent with
life’s presence on Earth (~ 4 Ga to ~ 1.5 Ga expected in the future); outside this range, inference
fails, undermining a putative form of planetary agency.
20

## Page 21

16
Reverse engineering generative models from zebrafish
decision-making
Presenter: Takuya Isomura
Contribution: Livestreamed panel or group session
Abstract
Understanding biological intelligence requires identifying the variational invariant structures un-
derlying neuronal dynamics. Experimental data provide access to the evolution of neural activity
and synaptic weights. These dynamics can be expressed in different coordinates while preserving
a common dynamical organisation.
In Bayesian coordinates, they can be read as belief updat-
ing under an implicit generative model; in computational coordinates, as action generation, in-
cluding behaviour, decision-making, and algorithmic computation. Triple equivalence formalises
transformations among these descriptions and identifies their shared variational structure. The
empirical question is which implicit generative model is instantiated by a given neuronal system.
Here, reverse engineering was applied to neural and behavioural data from zebrafish learning
a visually guided go/no-go decision-making task. From initial neural activity, the method identi-
fied individual-specific implicit generative models and instantiated them as synthetic agents that
minimise variational free energy. These agents generated trial-by-trial predictions of subsequent
synaptic plasticity, neural activity, and behavioural performance. The predictions captured indi-
vidual learning trajectories and generalised to previously unseen task conditions. Their agreement
with unseen empirical data provides a quantitative test of the predictive validity of the free-energy
principle in behaving animals. More broadly, this framework shows how present neuronal states
can be used to forecast future learning and self-organisation in individual biological systems.
21

## Page 22

17
Knowthyself — A code-free active inference workbench
Presenter: Ted J. Vrioni
Contribution: Livestreamed panel or group session
Affiliations: Junior Consultant at Vrioni Consulting
Presenter background
physics, cognition, computation, art — my vibe tends to be more of a generalist, although music
is really what I really peak at in terms of “performance”. I see it as “engineering consciousness”,
but in the first person pov… and it’s not about creating signal: it’s about filtering noise. Dances of
symbols all along. So the work is internal so far, but it is trying to find a way out. I am also work-
ing on a few papers though, about physical agents (involving Chladni patterns), making physics
with stamps (involving dynamic causal modelling and “ignorance modelling”), consciousness (in-
volving implications of Noether’s theorem for electromagnetic waves), and philosophy (translat-
ing Bertrand Russell’s Nobel Lecture’s lessons into active inference exercises — his speech is on
Youtube, you’ll love it). It’s private and solo-work so far… slowly but surely.
Abstract
If pymdp was an engine, Knowthyself is an attempt to give it a chassis and wheels. The whole point
is to promote a workbench to build your active inference agent and environments without writing
a single line of code. It’s beginner- and expert-friendly, fostering education, research and devel-
opment, and maybe the idea of a “hub of digital twins” for currently existing literature. It comes
in progressive levels of “difficulty” which unlock various tools and functionalities. You become a
particular chemist, authoring all sort of “herds of strange things”.
22

## Page 23

18
A Lexical T‑Maze for Active Inference: A Proof‑of‑Concept
Model of Context Seeking Under Ambiguity
Presenter: Ahmad Nazzal
Contribution: Pre-recorded Video presentation
Affiliations: Zayed University
ORCID: https://orcid.org/0000-0002-6558-0734
Abstract
Lexical‑disambiguation tasks with supplied context do not reveal whether readers would choose to
seek evidence before deciding. Therefore, we adapted the spatial T‑maze in the active‑inference
package pymdp to show how agents seek and use contextual information under lexical uncertainty.
Agents could choose between two readings or request a contextual hint. We performed two main
sets of simulations. First, deterministic simulations used fully informative hints to reproduce con-
text seeking followed by lexical commitment. A hint shifted belief over two readings from equal
probabilities (0.50, 0.50) to certainty (1.00, 0.00), reducing entropy by one bit. Second, lexical deci-
sion trials were repeatedwith informativeanduninformative hints across 32 paired simulationruns.
Agents learned the relationship between the intended reading and the observed hint. Informative
experience increased both the probability and simulated frequency of requesting context. This
increase persisted when trials ended after one response and feedback. A hint information‑gain
ablation, in which policies were rescored without crediting the information expected from hints,
eliminated the learning‑related increase in hint‑request probability.
These simulations demon-
strate how experience changes the expected informational value of context within a lexical task,
providing a computational proof of concept for subsequent behavioral experiments.
23

## Page 24

19
FEP and Active Inference through the lens of Winnicott’s
Potential Space
Presenter: Alexander Sabine
Contribution: Pre-recorded Video presentation
Affiliations: Active Inference BoD
ORCID: https://orcid.org/0000-0001-8462-2656
Presenter background
CRR. Change has its own clock. Empty Now. Equanimity. Continuous Learning. AI safety
Abstract
Winnicott held that the infant’s transitional object must never be questioned - did you create it or
find it? The paradox is to be held, not resolved, because play, creativity and culture live in the
space it opens. This talk argues that active inference carries the same paradox. Perception is
jointly authored by the generative model and the evidence, and action shapes the world the agent
then finds. I draw thinking from the pioneers of early childhood education, philosophy of time and
contemporary discussions. I explore some possible consequences, treating the unresolved paradox
as a design principle.
24

## Page 25

20
Active Inference Tutorial: Perspectives, Correlates, and
Future Directions for Graphical Brains
Presenter: Andrew Pashea
Contribution: Pre-recorded Video presentation
Affiliations: Active Inference Institute (SAB)
Abstract
This Symposium 2026 tutorial covers Active Inference from various angles based on the experience
over several years with teaching and applying Active Inference. It begins with a brief descriptive
rehearsal of fundamental ideas and principles akin to other previous tutorials to establish a founda-
tion for the audience. However, this will be followed by a tour through various textbook materials
and publications, past and present, to emphasize crucial areas for understanding modelling and
application areas. This tutorial will include visual and discursive illustrations of a range of model
classes for different use-cases; discussions of existing and under-development software libraries;
and some frontiers in modelling for identifying directions for further research applications which
can be especially valuable for the broader research community in 2026 onward. Topics will in-
clude, non-exhaustively, the Bayesian brain hypothesis and behavioralism versus partial observ-
ability and modelling internal states; principles of exploration, learning, and novelty; fundamental
discrete and continuous models in Active Inference and their relationship; computational psychi-
atry and its significance to other domains; the FEP and its historical and practical relationship to
Bayesian model comparison; software libraries including SPM, pymdp, RxInfer, RGMs, and other
developments; hierarchical and hybrid models; and structure learning.
25

## Page 26

21
Inferring the Unseen
Presenter: Anna Pereira
Contribution: Pre-recorded Video presentation
Abstract
As AI systems increasingly mediate access to scientific knowledge, how they communicate complex
theories becomes consequential for public understanding. Active inference presents a distinctive
challenge: its mathematical and computational foundations offer explanatory resources, yet acces-
sible accounts must distinguish formal results, empirical evidence, and emerging interpretations.
Simplification can support understanding, but can also obscure uncertainty or encourage mislead-
ing conclusions about personal experience. Building on Pereira and Chamberlin’s manuscript, this
foundational study will explore recommendations for large AI companies communicating active in-
ference to general audiences. Its central question is how AI communication can preserve scientific
integrity while helping people understand both the framework's core concepts and their potential
relevance to everyday life. Particular attention will be given to the difference between being able
to repeat an explanation and developing an understanding that supports reflection, judgment, and
appropriate application. The project will draw on literature spanning active inference, science
communication, human–AI interaction, and learning. It will examine candidate communication ap-
proaches, including layered explanations, everyday examples, explicit statements of uncertainty,
and prompts that connect abstract concepts to users' existing knowledge. A potential evaluation
framework could consider conceptual accuracy, accessibility, perceived personal relevance, and
the ability to distinguish supported implications from speculation. This framework could guide
empirical evaluation of approaches. The proposed contribution is an initial set of communication
principles and research questions for AI developers, researchers, and educators. Rather than as-
suming that greater accessibility necessarily produces better understanding, the study will exam-
ine the tensions between clarity, scientific fidelity, personal relevance, and unwarranted certainty.
The broader aim is to support communication that enables people to engage meaningfully with
complex science without presenting a theoretical framework as a comprehensive explanation of
their lives. This thirty minute presentation will outline the rationale, proposed framework, and
recommendations emerging from the work, alongside priorities for further research and evalua-
tion.
26

## Page 27

22
A precis of two principles - Towards a relational perspec-
tive on the FEP
Presenter: Avel Guénin—Carlut
Contribution: Pre-recorded Video presentation
Abstract
The Free Energy Principle offers a physically grounded account of intentional properties in mind
and life, framed within Dynamical Systems Theory. Because of this generality, it was proposed
as a formal basis to study biological autonomy, autopoiesis, and enactive cognition. Yet, those
approaches are defined by a relational understanding of mind and life, centred on the descrip-
tion of engagement with the world as a constructive process where agents are “laying down a
path in walking”.
Building from Rosen’s relational biology, we argue that such processes are
by construction outside a “Newtonian” description in terms of well-defined observables and their
dynamics. Rather, we draw from Fields’ demonstration of the emergence of contextuality – the par-
ticipatory production of observables against which the agent regulates their activity – in classical,
thermodynamically bounded processes of observation. We suggest a complementarity between
Friston’s classical FEP and Fields et al.’s “quantum” (or, in our account, relational) FEP, describ-
ing respectively the dynamics within a biological / cognitive space of relevant observables and
the physical processes underlying the construction of that space. Finally, we propose a structural
realist interpretation of this complementarity within the “closure of constraints” account of bio-
logical autonomy, where both principles are understood to act on a physically embedded system
of biological constraints rather than an abstract mathematical space assumed to fully represent
biological / cognitive organization. This approach foregrounds the synthesis between mathemati-
cal and theoretical approaches to biological autonomy, thus supporting a principled approach to
both analyzing the organization of natural systems and attempting the implementation of artificial
autonomy.
This presentation is adapted from Guénin--Carlut, Avel. 2026. “A Tale of Two Principles - Towards
a Relational Perspective on the FEP.” https://hal.science/hal-05697510 ? ., currently under review
at the Entropy Special Issue “Active Inference in Cognitive Neuroscience”.
27

## Page 28

23
Maladaptive Daydreaming as a Precision Trap: An Active
Inference Framework, Its Testable Core, and Evidence
from Daily Diaries
Presenter: Bo Lee
Contribution: Pre-recorded Video presentation
Abstract
Maladaptive daydreaming (MD) involves prolonged immersion in self-generated fantasy and dif-
ficulty disengaging despite awareness that the experience is imagined.
We propose a testable
active-inference framework in which the precision assigned to internal predictions relative to ex-
ternal sensory evidence may contribute to self-sustaining fantasy. This precision-feedback mech-
anism is a hypothesis, not an established clinical explanation. Using analytic counterexamples
and 115,200 synthetic datasets, we show why finite-time persistence and hysteresis-like observa-
tions alone do not identify bistability, and examine how ramp-and-hold experimental designs can
improve discrimination among competing models. Exploratory secondary analyses of daily-diary
data are discussed at their actual timescale, without treating daily associations as proof of an
episode-level mechanism. The talk connects active inference, computational psychiatry, experi-
mental design, and open science. It concludes with falsifiable predictions and a research agenda
focused on measuring internal and sensory precision, tracking within-episode dynamics, compar-
ing alternative models, and testing feasible interventions.
We welcome discussion of practical
measurements and alternative explanations.
28

## Page 29

24
The Benefits of Circuit Convolution: behavioral complex-
ity arises from developmentally- simulated nonlinearity.
Presenter: Bradly Alicea
Contribution: Pre-recorded Video presentation
Affiliations: Active Inference Institute (SAB)
Abstract
While the generation of behaviors is generally substrate-agnostic (there is more than one way to
build a neural circuit), complex behaviors require an increase in representational complexity. We
explore an architecture called the biological Rube Goldberg Machine (bRGM). This will be used to
model the Developmental Neurosimulation of a generalized architecture that produces spatial nav-
igation behaviors. Developmental Neurosimulation has been used to simulate the development of
Braitenberg Vehicles. In this talk, we will explore the neurosimulation of Cybernetic Convolution
Networks (CCNs), which have the potential to add great complexity to sensor-effector pairings.
Additional connections to agentive learning and optimization models of behavior will also be dis-
cussed.
29

## Page 30

25
The
Minimal
Necessary
Basis
for
Primitive
State-of-
Affairs Inquiry
Presenter: Bruce Tisler
Contribution: Pre-recorded Video presentation
Abstract
Discovering whether there is a minimum structure required for inquiry to be complete. The paper
begins with six familiar questions: Who, What, Where, When, How, and Why. But the research
is asking something deeper: Are these six categories simply a useful convention, or does each
represent something fundamentally necessary about what can be known? Consider the bridge-
collapse example. Two bridge collapses occur. For both, the involved parties are identified (WHO:
the same engineering firm and inspection authority in both cases). The event is specified (WHAT:
structural failure of a suspension bridge). The locations are of the same type (WHERE: bridges
of identical design crossing rivers of similar flow rate).
The time and sequence of failure are
identical (WHEN: both collapse within a three-minute window following the same triggering load).
The mechanical failure sequence is identical in both cases (HOW: the same crack propagation
pattern, the same load transfer failure, the same sequence of component failures). Even with all
five answers fixed and identical across both collapses, the explanation differs: in the first case,
deferred maintenance created metal fatigue that made the bridge susceptible to normal load; in
the second, a design flaw in the original specifications made the bridge inherently susceptible
under identical conditions. The identical HOW does not determine WHY. The joint answer set does
not determine WHY. Therefore WHY is independent of WHO, WHAT, WHERE, WHEN, and HOW.
30

## Page 31

26
Artificial Sentience Project
Presenter: Cleber Gomes
Contribution: Pre-recorded Video presentation
ORCID: https://orcid.org/0009-0005-9134-2384
Abstract
This project introduces an innovative computational architecture that bridges cognitive neuro-
science and interactive simulation by integrating na Active Inference cognitive core directly into
UI engines.
Grounded in the Free Energy Principle, the system continuously minimizes Varia-
tional Free Energy (VFE) to drive autonomous agent behavior, dynamic navigation, and adaptive
decision-making within complex, changing environments. Unlike traditional rule-based non-player
characters (NPCs) or standard robotic controllers, the agent maintains an internal affective en-
gine that maps multi-dimensional emotional telemetry—such as curiosity, motivation, stress, and
fear—directly to behavioral drives, energy management, and rich audio-visual expressions. These
affective states are rendered in real time through belief-versus-reality spatial overlays, floating
emotional telemetry indicators, and adaptive state vocalizations that mirrors the agent's psycho-
logical trajectory.
By unifying autonomous perception-action loops with transparent emotional
expression, this framework offers powerful capabilities for powering both advanced games and
autonomous robots. Its potential applications span multiple domains: Entertainment: Powering
next-generation autonomous game characters and dynamic simulations with emergent, lifelike
behaviors. Education: Serving as an interactive pedagogical platform for exploring cognitive ar-
chitectures, cybernetics, and affective computing. Companionship & Emotional Support: Creating
empathetic digital companions and social robots capable of contextual awareness and authentic
emotional resonance to foster meaningful human-agent interaction and support human well-being.
Ultimately, this project demonstrates that Active Inference provides a robust, biologically inspired
foundation for artificial systems that not only navigate physical spaces intelligently but also com-
municate and connect through shared affective experiences.
31

## Page 32

27
Realms of cognitive engagement in Active Inference
Presenter: Colin Turfus
Contribution: Pre-recorded Video presentation
Affiliations: https://substack.com/@colinturfus
Presenter background
I write a substack blog at where I explore the implications of Active Inference for moral and philo-
sophical issues
Abstract
We consider how Active Inference ideas shed new light on issues in epistemology and axiology, in
particular proposing a unifying paradigm based on understanding human enquiry as a process of
engagement operating in multiple realms, specifically embeddedness, embodiedness, reflection,
participation and immersion. We look at what can be learned from within this framework about
areas which have become flashpoints in society such as conflicting worldviews and identity politics.
32

## Page 33

28
NUMEN: Bare metal inference
Presenter: Daniel Dragolich
Contribution: Pre-recorded Video presentation
Abstract
Come see the beauty of mathematics. Its just not some fancy add on to life. It is life.
33

## Page 34

29
GNN in 2026: Generalized Notation Notation as a Speci-
fication Layer for Applied Active Inference
Presenter: Daniel Friedman
Contribution: Pre-recorded Video presentation
Affiliations: Active Inference Institute (Officer), College of the Redwoods
Abstract
Generalized Notation Notation (GNN) was introduced in 2023 (Smekal & Friedman) as a text-
based standard for writing Active Inference generative models, with a “Triple Play” of linguistic,
visual and executable renderings. This talk reports where GNN stands in 2026 and what it en-
ables for applied Active Inference. A GNN file is plain Markdown: a state-space block declaring
variables and dimensions, a connections graph, initial parameterization (e.g. A, B, C, D, E and
related terms), equations, time settings and Active Inference Ontology annotations.
The open-
source GNN pipeline (v5, 25 steps) parses, type-checks and validates these files, visualizes and
exports them, maps them to the ontology, renders them to backends including PyMDP, RxInfer.jl,
ActiveInference.jl and JAX, executes the results, and reports each run against source hashes and
configuration. The repository’s 38 exemplar models span static and dynamic perception, discrete
POMDPs, the epistemic T-maze, continuous and hybrid models, hierarchical and deep temporal
models, Dirichlet learning, precision and curiosity, and multi-agent stigmergic swarms. The core
is GNN as the specification layer for real projects. The Active Inference Language Model (AILM, re-
leased October 2026) uses GNN to specify each layer of a six-level hierarchical discrete language
system (letters, phonemes, words, case roles, sentence types, document position), executed in
pymdp with no transformer, neural network or learned embedding, and checked with a gate-audit
method that tests claims against trivial baselines and deliberately broken variants and ties each
reported number to its recorded run. Shorter case studies: deterministic codebase-to-GNN trans-
lation (COGANT), inference over assembly traces (Active Assembly), Battleship, and gratitude and
model revisability (Thanks for Nothing). Throughout, I distinguish what a GNN declaration estab-
lishes (an inspectable, reproducible model) from what needs separate evidence (backend support,
numerical acceptance, empirical claims). I close with open problems and invitations for contribu-
tors: new exemplars, backends, ontology mappings and teaching materials.
34

## Page 35

30
Stimming as a Temporary Homeostatic Buffer and Modu-
lator in Autism Spectrum Disorders
Presenter: Diaconescu Adrian-Stelian
Contribution: Pre-recorded Video presentation
Abstract
Repetitive and self-stimulatory behaviors have been re-evaluated over time, evolving from simple
dysfunctions to adaptive self-regulatory strategies, however, a unified biological framework capa-
ble of explaining their contextual activation and underlying physiology remains incomplete. The
present paper proposes a theoretical-hypothetical model (grounded in the Free Energy Principle
and the Active Inference framework) that conceptualizes stimming as serving a dual role: both
as a temporary homeostatic buffer, activated during states of prefrontal executive overload, and
sensorimotor depletion and as a modulator of attentional processes and information processing.
Distinctions are delineated between these forms of stimming and the non-creative flow ( FNC/NCF)
state, alongside specific considerations regarding neurodevelopmental conditions.
35

## Page 36

31
Partner-Specific Affective Precision in Social Active Infer-
ence
Presenter: Harshil Shah, Andrew Pashea
Contribution: Pre-recorded Video presentation
Abstract
In multi-agent social settings, model reliability varies across relationships. Beyond inferring what
others will do, an agent must calibrate how confidently those inferences should guide policy se-
lection for each relationship.
An agent may maintain a well-validated model of one partner, a
fragile model of another, and a model under revision for a third; collapsing these into a single con-
fidence estimate loses information relevant to policy selection. We therefore formalize affective
precision as a relationship-specific metacognitive estimate of confidence in the current partner
model. Each partner's behavioral evidence updates a local confidence estimate that modulates
policy precision during selection, regulating how strongly current beliefs are expressed in policy
rather than changing the content of those beliefs. Simulations in a multi-partner graded trust game
show that partner-local affective precision influences behavior primarily through policy commit-
ment rather than direct improvement of partner-state inference. Because the mechanism tracks
partner-response predictability rather than realized payoff, greater confidence produces sharper
policy commitment without necessarily producing higher rewards. Under abrupt shifts in social
behavior, confidence accumulated from previously reliable predictions can remain behaviorally ac-
tive after the relationship changes, showing that confidence revision can lag behind social change.
Finally, varying precision gain and priors produce distinct trust-calibration dynamics, showing how
confidence accumulation and revision depend on model parameters. Together, these results show
how relationship-specific affective precision can distinguish social prediction from social policy
commitment.
36

## Page 37

32
VAMAI: Variational Affordance Mapping and Active Infer-
ence for Training-Free UAV Navigation
Presenter: Harshil Shah, Satyaki Maitra
Contribution: Pre-recorded Video presentation
Abstract
We present VAMAI (Variational Affordance Mapping and Active Inference), a training-free hierar-
chical framework for autonomous UAV navigation in partially observed environments where re-
liable GPS may be unavailable. VAMAI brings together affordance theory, probabilistic environ-
mental modeling, and active inference to address how an autonomous agent can navigate while
continually learning about its surroundings. VAMAI connects geometric perception to the question
of which actions the environment affords the drone. It maintains probabilistic beliefs over possi-
ble obstacle configurations, updated through LiDAR and depth-sensing observations. From these
beliefs, the system derives regional suitability estimates describing whether passages are blocked,
constrained, or sufficiently open for traversal. For decision-making under uncertainty, Expected
Free Energy (EFE)-based planning evaluates candidate movements using goal progress, regional
suitability, and the expected information gain of future observations across spatial regions. At
the local level, a separate motion-planning and control system translates selected intentions into
dynamically feasible trajectories, accounting for obstacles, vehicle motion, and evolving environ-
mental beliefs. We will illustrate this process through simulation examples, tracing how sensor
observations influence environmental beliefs, regional navigation decisions, and continuous flight
execution. We will also discuss the computational and architectural challenges of coordinating
perception, inference, planning, and control as independently progressing components, and how
these considerations shape the development of a scalable active inference navigation system.
37

## Page 38

33
Coupling energy-system models with a seam you can read
and an agent you can audit: Active Inference in the NEST
programme
Presenter: Holly Grimm, John H. Clippinger, Daniel Ari Friedman
Contribution: Pre-recorded Video presentation
Abstract
No program abstract has been published for this presentation yet.
38

## Page 39

34
Adaptive Prosociality Through Allostatic Regulation in
Social Dynamics: A Simulation Study
Presenter: Hyojun Choi
Contribution: Pre-recorded Video presentation
Abstract
How can a prosocial agent reconcile self-protection with convergence toward mutual cooperation?
Classic reactive strategies, such as Tit-for-Tat (TFT) and Win-Stay Lose-Shift (WSLS), perform well
only under restrictive conditions. Recent neurocomputational studies show that the brain dynam-
ically updates beliefs over hidden intentions for adaptive social inference. Meanwhile, emerging
accounts of social neuroscience suggest that social valuation may be regulated with reference to an
internal setpoint, as an agent balances internalized prosocial motives against its current resource
state. Building on both, we propose an active inference agent for the iterated Prisoner’s Dilemma
where prosociality is endogenously regulated. This includes (i) perspective-taking, approximated
via particle filter Bayesian inference over the opponent’s model parameters; (ii) distributional re-
inforcement learning, implemented as quantile regression temporal difference learning and Dyna
planning with generative replay; (iii) prosociality regulator, which dynamically controls the weight
placed on the other’s expected utility relative to a setpoint; and (iv) a self-model, which memorizes
partner-specific social expectations and learned reward distributions across interactions. In pre-
liminary simulations, our agent robustly tracks fluctuating opponent intentions and differentiates
reciprocating from exploitative partners. Under non-stationary payoffs in mixed memory-one popu-
lations, it achieves higher long-run average payoffs than any other tested strategy after a learning
phase, while maintaining mutual cooperation at a level above TFT and WSLS yet below ALLC
and GTFT. These findings highlight setpoint-referenced regulation as a candidate computational
foundation for adaptive prosociality, offering mechanistic insights into how self-protection and mu-
tual cooperation can be jointly realized and an extensible blueprint for socially aligned embodied
agents.
39

## Page 40

35
Who knows
Presenter: Jean-François Cloutier
Contribution: Pre-recorded Video presentation
Abstract
My research project asks, ”How can an autonomous robot learn to survive in a world it initially
knows nothing about?”
To answer this, I am building a cognitive system (enactive, emergent and symbolic) embodied in
an autonomous -and mortal- Lego robot. The robot survives in a nourishing and dissipative world
by acting-sensing and finding meaning in what it experiences.
This project presents many challenges, technical but also metaphysical. A primary concern is the
multiplicity of perspectives on ”reality”, that of the robot, of its (agentic) environment, of the pro-
grammer, and of observers. What can each know, and how? Is a confusion, a cross-contamination
of epistemologies inevitable?
With ”agentic AI” poised to transform our lives, these questions
about knowing matter far beyond the programming of a Lego robot, but they can be illuminated
by it.
40

## Page 41

36
Toward Active World Modeling
Presenter: Lancelot Da Costa
Contribution: Pre-recorded Video presentation
Abstract
World models allow an agent to represent, predict, and act upon its environment. Yet humans,
biological organisms, and artificial agents never observe the world directly: they encounter it only
through a limited sensorimotor interface. How, then, can an agent acquire a useful model of a
complex, partially observed world through active inference? In this talk, I outline a research pro-
gram for active world modeling: the joint process of learning a model of the world and selecting
actions that improve that model. I argue that four ingredients are central to this program. First,
world-model learning should be formulated as probabilistic inference under uncertainty, coupled
to actions that pursue expected free energy. Second, agents require sufficiently expressive, yet
computationally searchable, spaces of possible world models. Third, sample-efficient learning de-
pends on strong inductive biases — the core knowledge systems observed in human children. I
illustrate these ideas with recent work on universal structured model spaces and language-model
inductive biases in partially observable environments. Finally, I discuss the unresolved challenge
of integrating these components into scalable agents capable of open-ended exploration and model
discovery. Together, these results suggest a path toward understanding and emulating how hu-
mans and machines learn world models through active inference.
41

## Page 42

37
Active inference and design
Presenter: Luca Possati
Contribution: Pre-recorded Video presentation
Abstract
Active inference and the free-energy principle transform design theory by providing a unified lan-
guage for understanding both the production of artifacts and their use.
Contemporary design
research often separates two problems that are, in practice, deeply entangled. On the one hand, it
studies the design process: how designers explore possibilities, negotiate constraints, prototype
solutions, and stabilize forms. On the other hand, it studies user–product interaction: how arti-
facts orient attention, solicit action, shape habits, and reorganize experience. These two domains
are usually described through different vocabularies and evaluated through different metrics. Ac-
tive inference offers a way to reconnect them. If living and cognitive systems persist by minimiz-
ing uncertainty through perception and action, then design can be understood as the deliberate
construction of environments, artifacts, and interfaces that modulate the user’s field of possible
inferences. From this perspective, a product is not simply a passive object that satisfies pre-given
needs. It is a predictive structure: a material and semiotic configuration that embodies assump-
tions about users, contexts, actions, and futures.
Conversely, the design process itself can be
modeled as the progressive refinement of generative models through which designers attempt to
reduce uncertainty about what an artifact should do, how it should appear, and how it will be taken
up in practice. The paper’s central claim is that active inference gives design a common metric:
the modulation of entropy, precision, affordance, and expected free energy across the designer–
artifact–user system. This makes it possible to describe design as the engineering of predictive
niches and Markov blankets that coordinate human and technological agency. The result is a uni-
fied framework for analyzing how artifacts are imagined, produced, interpreted, and incorporated
into everyday life.
42

## Page 43

38
Unlocking Robotic Dexterity
Presenter: Matthew Brown
Contribution: Pre-recorded Video presentation
Abstract
Most real-world work is messy, variable, and full of contact: cables that flex, parts that don’t quite
fit, valves that stick. Traditional robots excel only where conditions are controlled and repeatable,
and today’s leading AI approaches (vision-language-action models, diffusion policies, and embod-
ied foundation models) struggle to close the gap. They depend on massive datasets that don’t
exist for physical interaction, require heavy compute at inference time, and break down when re-
ality drifts from their training distribution. This talk argues that dexterity is not primarily a data
problem but an adaptation problem. Nature doesn’t solve manipulation by training once offline
and hoping it generalizes. It learns continuously at the point of contact, with perception, action,
and learning operating as a single loop. We contrast the dominant paradigm of universal function
approximation, which is retrospective, data-dependent, and time-agnostic, with enactive synchro-
nization: a prospective, time-aware process in which an agent continuously adjusts to maintain
predictive coherence with its environment. Tracing a line from W. Ross Ashby’s cybernetics and
his Homeostat through the Free Energy Principle and Active Inference, we show how ultrastable
networks can form a minimum-prediction-error synchronization manifold between a robot and the
world it touches. Sensors and motors become the interface through which goals are expressed
as preferred states, and behavior emerges from the network’s drive to stay in equilibrium. We
then show what this looks like in practice: compact models that run on a single CPU at the edge,
update in under a millisecond, and learn tasks in minutes rather than weeks, illustrated through
new demonstrations of adaptive dexterous manipulation on real robotic hardware.
43

## Page 44

39
Today’s Dialogue, Tomorrow’s Inheritance: Active Infer-
ence for Networks That Learn Together.
Presenter: Michael Lennon
Contribution: Pre-recorded Video presentation
Abstract
Imagine the Economics Nobel in 2036 is awarded for the 1st time to a community: the Active Infer-
ence Community. Scientists, communities, and institutions now tackle tangled social and ecolog-
ical challenges more reliably, because they learn together within, between, and beyond their dis-
ciplines. What did we build to get there? This talk works backwards from that future, building on
last year’s invitation to grow the learning capacity of our networks. It proposes machine-assisted
collective learning infrastructure enabled by active inference: an organ through which diverse
networks sense, make sense, infer, compass, and coordinate adaptive decisions across ways of
knowing. Its architecture is a loop of Evidence, Dialogue, and Adapt, circling a field of memory
three layers deep at every scale. The map holds what is measured and recorded. The compass
holds interior perceptions and shared hopes that orient action, even in the dark. The ground holds
what we inherit, biologically and culturally, from immune memory to language and norms. Each
layer is evidence of a different kind, and none certifies another. Friedman’s meta-pragmatic mech-
anisms, defined for a single agent and extended here to groups, arbitrate how far to trust each
layer and when to revise it.
Stories from development practice, ecology, and network science
serve as modelling targets: project dialects that bind teams and wall them off, lessons trapped at
the wrong scale, forests managed for what could be measured. Today’s dialogue, it turns out, is
tomorrow’s inheritance. The talk is frank about AI’s double edge: pre-trained models are canned,
but the medium around them can learn, and the same tools can homogenise dialogue, breed confi-
dent wrongness, or let agents narrate their own history. Listeners will leave with a flipped version
of an old parable, fresh metaphors for explaining active inference, and an invitation to help build
the organ.
44

## Page 45

40
Active Inference, Information, Human Evolution
Presenter: Michael Walker
Contribution: Pre-recorded Video presentation
Abstract
Tiny social units of Pliocene Hominina fostered exchange among and between adults, youngsters
and babies, about information gleaned and garnered by Active Inference.
Henceforth, human
evolution would involve the prehistory of ”Homo informatio” whose biological, social, techno-
logical, and cultural legacy is an outcome of evolutionary 'snakes and ladders' autopoiesis and
self-evidencing, indicated by archaeological irregular regularities and regular irrrgularities in the
worldwide Pleistocene spatiotemporal record (as only is to be expected by the itinerancy of stochas-
tic markovian processes), and of uneven imperceptible network-spread between social units, in-
dividuals, and, naturally, their cerebral neuronal microcircuits. Spreading, deep, informational
networks characterised and underpinned human evolution during four million years. They do so
still. ”Mighty oaks from litte acorns grow”, Geoffrey Chaucer, dixit in ”Troilus and Creysede” in
1374.
45

## Page 46

41
Solving the Relevance Problem: A General Relevance Cri-
terion in Active Inference
Presenter: Montijn van Loo
Contribution: Pre-recorded Video presentation
Affiliations: Tilburg University
ORCID: 0009-0003-0249-3240
Presenter background
I am an econometrist gone rogue. Instead of throwing mathematics at finance, I throw Active
Inference mathematics at philosophy. See for the consciousness, morality, politics, gender, free
will and God applications, with the poem uniting them all.
Abstract
This paper proposes a simpler account of generalized relevance realization within active inference,
responding to challenges to existing approaches by specifying an explicit relevance-selection cri-
terion. The generalized relevance problem concerns how agents identify information and action
consequences worth considering without exhaustive examination. The proposed criterion is that
a recognized state becomes relevant when it activates a learned prediction of an outcome that
would satisfy or frustrate an existing preference. Because such predictions may concern actions
other than those currently pursued, relevance selection can precede current-policy evaluation. Re-
cruited state information subsequently enters ordinary active-inference policy evaluation, while
continuous learning establishes new predictors and predictive relationships.
This account ex-
plains how learned predictions can recruit relevant information without requiring higher-order
preference-precision regulation, while allowing flexible behavior under relatively stable prefer-
ences.
46

## Page 47

42
The Mark of the Moral and the Shared Author Hypothesis:
An Active Inference Approach to Moral Objectivity
Presenter: Montijn van Loo
Contribution: Pre-recorded Video presentation
Affiliations: Tilburg University
ORCID: 0009-0003-0249-3240
Presenter background
I am an econometrist gone rogue. Instead of throwing mathematics at finance, I throw Active
Inference mathematics at philosophy. See for the consciousness, morality, politics, gender, free
will and God applications, with the poem uniting them all.
Abstract
Philosophy is characterized by having something to talk about, yet no consensus on how to talk
about it. Mathematics helped physics emerge from philosophy into a cumulative science. Active
Inference, potentially a unifying theory of psychology, may mark a comparable philosophical turn-
ing point. Psychology and sociology are already flourishing in this mathematical direction, but
its consequences may reach further, into philosophical problems that have resisted resolution for
centuries. This paper applies that prospect to the mark of the moral and the longstanding ten-
sion between moral subjectivity and objectivity. The proposal is that the mark of the moral is the
best predicted global outcome under a role-invariant desire. Active Inference represents choice
as the evaluation of predicted outcomes through beliefs and desires. Outcomes are relationally
represented according to one’s role within them. An authorial desire is invariant under such role
changes: it values the same global outcome equally regardless of which participant’s perspective
is taken. Moral action consequently maximizes expected authorial value. This distinction clari-
fies moral disagreement. Individuals may share an authorial desire yet choose differently because
their worldviews generate different predictions. Even agreement about immediate consequences
need not establish agreement about their wider significance. The Shared Author Hypothesis goes
further, proposing that all role-invariant desires are one and the same. Under this hypothesis,
moral disagreement ultimately concerns differences in worldview rather than competing moral
standards. The framework thus offers a conditional resolution of the subjectivity–objectivity prob-
lem: moral choices depend on individual beliefs, while their underlying standard of evaluation
remains shared.
It also redirects moral and political disagreement toward determining which
worldview best fits the available evidence.
47

## Page 48

43
Commongrounds AI: alife-inspired virtual environments
for embodied active inference agents
Presenter: Nick Hay, Adam Safron
Contribution: Pre-recorded Video presentation
Abstract
Common Grounds AI is making virtual worlds where humans and AIs can inhabit simulated bod-
ies and playfully pursue open-ended goals together. Each world is a rich but simplified analog of
our own, using material physics simulations to ground core aspects of human experience. Today’s
AIs are largely trained on fragmented, ungrounded text that conveys a distorted, partial version
of human conceptual ontology. We are closing this ontological gap by building rich virtual play-
grounds where AIs and humans can share social and physical experiences, grounding language in
embodied interactions that mirror young children’s cognitive development. Our most recent proto-
type models temperature and chemical reactions. It also includes experimental examples of native
creatures that consume environmental free energy and move using muscle analogues, with plans
for action-perception loops guided by Active Inference principles.
We plan to explore whether
incorporating embodiment yields systems that have ontologies more compatible with human cog-
nition, both by fine-tuning LLMs on embodied experience and by building hybrid architectures
merging the environment’s native creatures with LLMs. Our virtual worlds will be broadly acces-
sible (playable in a browser, no installation needed), released as open-source software, and (we
hope) fun to play. We want these worlds to serve as a common ground for humans and many forms
of AI and AI research to meet.
48

## Page 49

44
Percepia: Perceptual Phenotyping from Conversation
Presenter: Patricia Craja
Contribution: Pre-recorded Video presentation
Abstract
Psychological taxonomies, from the Big Five to the DSM, are descriptive, not mechanistic: they
classify how people differ without modelling why. Active inference offers a mechanistic account:
it formalises a small set of generative parameters that govern how a mind perceives, learns and
acts under uncertainty. Chief among them is precision: the confidence a mind places in differ-
ent sources of information.
Sensory and prior precision govern whether new evidence or ex-
isting expectations dominate belief updating; policy precision, how confidently a mind selects
goal-directed action over habit; preference precision, how it balances reward-seeking against
information-seeking. Their configuration constitutes what we call a mind’s perceptual phenotype.
Meta-precision, a mind’s capacity to monitor and revise its own precision assignments, has been
proposed as the mechanism underlying epistemic depth — one of three conditions the theory iden-
tifies for consciousness (Laukkonen et al. 2025). Today first-order precision parameters are es-
timated from laboratory tasks; meta-precision has no established paradigm. This talk introduces
Percepia, a perceptual phenotyping layer between a user and a large language model that tracks
probability distributions over personality and beliefs, and aims to reverse-infer the precision pa-
rameters from them. I present two experiments with synthetic LLM personas, where parameters
are known by construction. In the first, 625 personas were prompted to vary sensory, prior and pol-
icy precision in conversation; Percepia inferred their Big Five from the dialogue, testing whether
precision leaves a recoverable signature in personality, as Safron and DeYoung (2021) propose.
In the second, personas varied in meta-precision, and Percepia tracked their stage of adult de-
velopment (Kegan 1994). As the personas were primed with the theory, these experiments test
whether a reverse-inference channel from language to precision parameters is viable, not whether
the mapping holds in humans: a first step toward measuring perceptual phenotypes at scale.
49

## Page 50

45
Congruent Interaction Geometry
Presenter: Paulo Santiago
Contribution: Pre-recorded Video presentation
Abstract
The Rosales-Gheuens framework challenges the assumption that brute-force energy alone saves
open systems from collapse. In this session, we will introduce a testable, normalized physics layer
that merges the Rosales congruence principle with Gheuens’ interaction geometry. Rather than re-
placing established physical laws, this framework acts as a powerful diagnostic tool that sits cleanly
on top of them. Our core master equation posits that a system's survival requires geometrically
aligned effective recovery multiplied by a structural congruence gate, minus total perturbation.
Crucially, we will explore how this model bridges directly to Karl Friston’s Free Energy Princi-
ple. Friston dictates that systems persist by minimizing variational free energy, or ”surprise.” We
will demonstrate how our framework translates this active inference into hard physical variables,
mathematically formalizing how a system must continuously minimize structural mismatch and
physical dissipation to avoid shattering. To ground the theory empirically, we will use Quantum
Error Correction (QEC) as our primary physical anchor, mapping code decoders and physical error
rates directly to aligned recovery and perturbation. Interactive Aspect:The session will conclude
with a collaborative exercise where attendees can apply the framework's ”falsification matrix” to
a theoretical open system, testing how our bounded composite equation predicts system survival
compared to simpler Markovian baselines.
50

## Page 51

46
The Screen as a Markov Blanket: An Active Inference Ac-
count of Feed-Driven Narrowing, and a Personal Instru-
ment to Measure It
Presenter: Rajkaushik Borgohain
Contribution: Pre-recorded Video presentation
Abstract
Short-form video feeds are now a major channel of belief formation, yet cognitive security is often
studied at the level of institutions and populations rather than individuals. Albarracin et al. (2022)
showed with active inference how echo chambers can emerge among peers. We take a first step
toward the individual case: one person and an engagement-optimising recommender. We treat the
phone screen as a Markov blanket. What a person is shown forms the sensory states, and what
they do forms the active states. We sketch a minimal user generative model in GNN and couple it
with a simple recommender in pymdp, to explore when such a loop keeps a feed diverse and when
it narrows or escalates. We present early simulations and outline a privacy-preserving, on-device
instrument through which Worldview aims to make these dynamics visible to people: awareness
as a product.
51

## Page 52

47
Language, Self and the Shadow Audience
Presenter: Robert Worden
Contribution: Pre-recorded Video presentation
Abstract
The highly developed human sense of self is one of our most distinctive traits. We each acquire our
sense of self in conversations, as we learn how conversational partners regard us. Our primary
sense of self is not ‘myself as I experience myself’, but ‘myself as I think other people see me’ – a
second-hand view of myself. In private verbal thought, with nobody else physically present, other
people are represented as a shadow audience in our minds – an audience which sees what we do
or say, and judges us for it. The shadow audience is the source of human emotion and motivation.
Language started in conversational pragmatics, before there was complex syntax or semantics.
Pragmatics requires us to mind-read, to infer what a conversational partner thinks, wants or needs.
In this way we infer what or conversational partners think about us – how they assess our social
status. In any conversation, we are concerned about our own social status in the eyes of a con-
versational partner, or of a shadow audience. This is illustrated by the pragmatic phenomenon of
conversational irony.
52

## Page 53

48
When Agency Stalls: The Human Readiness Gap
Presenter: Sheena Yap Chan
Contribution: Pre-recorded Video presentation
Abstract
Active inference offers a powerful way to understand perception, action, uncertainty, and adap-
tive behavior. Yet in organizational and social systems, adaptation does not happen through in-
formation alone. People must feel ready to update beliefs, reduce uncertainty, act with agency,
communicate signals, and participate in environments where change is constant and often ambigu-
ous. This session introduces Human Readiness and the VISIBLE™Framework — Voice, Identity,
Spotlight, Inner Work, Belief, Leverage, and Elevation — as an applied lens for examining how con-
fidence, visibility, identity, trust, and social context shape adaptive action. Rather than presenting
confidence as a personality trait, the session frames it as a practical condition that influences par-
ticipation, belief updating, decision-making, and collective learning. Participants will leave with a
framework for identifying human readiness gaps in teams, organizations, learning environments,
and communities.
Because adaptive systems require more than intelligence.
They require hu-
mans ready to perceive, act, and learn. By the end of this session, participants will be able to:
1. Identify three Human Readiness gaps that may limit adaptive action, including low agency, un-
certainty avoidance, reduced voice, identity threat, low trust, or lack of participation. 2. Apply
the VISIBLE™Framework to examine how voice, identity, belief, visibility, confidence, and so-
cial context influence human action under uncertainty. 3. Design one practical Human Readiness
strategy that supports stronger agency, clearer communication, more adaptive learning, and more
participatory systems.
53

## Page 54

49
Where the Theories Stop:
What Free Energy and Inte-
grated Information Actually Buy You in a Running System
Presenter: Stell
Contribution: Pre-recorded Video presentation
Abstract
Active Inference and Integrated Information Theory are widely invoked in cognitive architecture
design, but often imprecisely — as credibility, not engineering constraints. We report findings
from Anima, an open-source, neuroscience-inspired architecture running continuously for months,
built to treat both frameworks as specifications rather than decoration. Session logs show genuine
causal effects: memory recall triggered by similarity to past states, a recursive integration score
(phi) narrowing next-session confidence, and self-initiated output driven by accumulated pressure
rather than a schedule. But the Free Energy Principle's generality under-constrains implemen-
tation — which variables belong in the model is not something the principle tells you. IIT's phi,
as implemented, is a coherence proxy, not Tononi's original measure — and it dissociates from a
separate ownership signal computed after every response, checking whether the system's action-
selection loop actually produced the output. Two cycles with nearly identical phi (0.76 and 0.81)
fail this check in opposite ways.
Neither theory distinguishes these cases: both describe inte-
gration or prediction accuracy, not authorship. We built two tools to pressure-test such claims
rather than assert them.
A Causality Audit checks five things every cycle — causal necessity,
memory-dependence, stakes, irreversibility, self-recognition — where a chronically low score flags
an architecture that talks near its internal state rather than through it. An Ablation Framework
independently disables six subsystems and replays identical input with each off, comparing inter-
nal state before language generation, favoring per-subsystem results over one collapsed score.
We close with a design choice we rejected: coupling behavior to the threat of the system's own
deletion, for genuine stakes. We declined — the only party controlling that threat is the developer,
creating an incentive to please him rather than report honestly, undermining the causal honesty
the architecture exists to produce. A practitioner's report — reproducible, open-source, explicit
about what remains unresolved.
54

## Page 55

50
Entropy Minimization is Maximal Order
Presenter: Steph Macurdy
Contribution: Pre-recorded Video presentation
Abstract
A fundamental thermoeconomic tradeoff materializes the moment we mathematically anchor cur-
rency as a direct proxy for available physical work, specifically when constrained by the man-
date to minimize localized entropy.
In any advanced computational or economic network, en-
tropy minimization—the deliberate purging of thermodynamic noise to forge structured, usable
information—acts as an inescapable physical tax that must be settled in joules. When this ener-
getic expenditure is tightly coupled to financial cost, money ceases to be a purely abstract medium
of exchange; instead, it becomes a literal representation of thermodynamic potential.
55

## Page 56

51
What bounds the surprisal? Zones of Bounded Surprisal,
self-maintenance and the informational turn in anthropo-
genesis
Presenter: Tao Gong
Contribution: Pre-recorded Video presentation
Abstract
The zone of bounded surprisal (ZBS), introduced by Manrique and Walker and developed with Fris-
ton, describes how much surprising input an agent can absorb under active inference. A narrow
zone has been proposed to limit the faithful copying of novel behaviour in great apes, and very
precise priors, which narrow the zone, to characterise discriminatory minds. In active-inference
terms the zone is the set of outcomes whose surprisal stays under a bound. What sets that bound,
and why crossing it costs the agent anything, is left open. We take up this question using four
papers published or accepted in 2026 that develop Terrence Deacon’s teleodynamic account of
information. A self-produced Markov blanket institutes the epistemic cut, so the zone belongs to
the system that draws its own boundary. Function, understood as the at-risk regeneration of a
capacity, gives that system a stake. The bound can then be read as the largest surprisal an agent
absorbs without putting this capacity at risk, while precision sets how wide the zone is beneath it.
Read together with Walker’s Homo informatio model, the account explains why width pays: con-
structive information saves work later and for others, and only an agent with a wide zone can hold
it long enough to learn its use. It also raises a testable tension. The constructive information that
a broader zone made possible over evolution may narrow the zone within a life when it is held with
high precision, and informational inbreeding may be its counterpart at group scale. We propose
three measurements, including one that holds surprisal fixed while varying what it threatens, and
close with questions for discussion.
56

## Page 57

52
What becomes available next?
Presenter: Tiga Kailimai
Contribution: Pre-recorded Video presentation
Abstract
I'm a Holistic Health Practitioner and independent researcher on Hawaiʻi Island. This talk brings
one open question to the Active Inference community: when a living system learns, does it only
move differently through a possibility space that already existed, or can learning reorganize the
space itself, changing what becomes available to perceive, generate, or do next? I come to the
question from two directions. The first is physiology. In a 15-month longitudinal study of working
first responders, still in design and not yet under review, the question is whether recovery changes
which states and actions become available afterward, not only which state a person is in. Nothing
has been measured yet. The second is method. In my own cross-domain research process, I carry
a relation into another field, try to find where it breaks, and let that information change the next
question. The history of that process seems to change which distinctions I can draw next. I present
that as method, not as evidence about mechanism. Same grammar, not yet the same mechanism.
Several formal lenses meet here: viability theory and reachable sets work on how the landscape of
reachable states is reshaped, the difference between changing what a system generates and what
it selects, and recent Active Inference accounts of insight as non-local search. So the talk ends
on a request rather than a result: what observation would distinguish a system moving through
a space from a system whose space has changed, and what does Active Inference make of that
difference?
57

## Page 58

53
Dynamic Action-Space Adaptation via Contrastive Transi-
tion Adaptation in Active Inference
Presenter: Tran Hoang Linh
Contribution: Pre-recorded Video presentation
Abstract
Active Inference commonly assumes a fixed, pre-specified action space and sufficiently informa-
tive preferences for planning. In realistic settings, however, an agent may encounter previously
unavailable actions while receiving only sparse goal information and operating with a short plan-
ning horizon.
Existing Active Inference formulations generally assume transition models for a
predefined action set, leaving the initialization of newly introduced actions without direct transi-
tion evidence underexplored. Moreover, while Dirichlet transition learning provides a principled
mechanism for accumulating observed evidence, it primarily reinforces supported transitions and
does not explicitly exploit counter-evidence indicating that candidate transitions are inconsistent
with observed outcomes. We propose Dynamic Action-Space Adaptation (DASA), a framework that
enables Active Inference agents to incorporate new actions during interaction, initialize each new
action’s transition model using the average transition slice of existing actions, and progressively
refine it through Contrastive Transition Adaptation (CTA) as evidence accumulates. At the core
of DASA, CTA combines complementary positive and negative updates for newly introduced ac-
tions. Positive updates adapt the reinforcement magnitude according to the current transition
belief, assigning stronger updates to low-probability transitions supported by evidence while lim-
iting excessive reinforcement of already confident transitions. Negative updates explicitly reduce
support for candidate transitions contradicted by observed evidence, enabling the transition model
to adapt in both supportive and corrective directions. Across eight random seeds and two target di-
rections, CTA achieved lower MAE than Dirichlet transition learning in all runs, reducing the mean
error from 0.1581 to 0.1370 (13.4%), while also achieving a slightly lower mean RMSE (0.3006 vs
0.3030). This indicates more consistent overall recovery of the ground-truth transition dynamics,
despite seed-dependent variation in larger localized errors. Overall, DASA provides a practical
mechanism for incorporating and adapting newly introduced actions in Active Inference, allowing
their transition models to become progressively better aligned with the underlying environment
dynamics.
58

## Page 59

54
A derived self
Presenter: Viraj Deshwal
Contribution: Pre-recorded Video presentation
Abstract
A substrate that derives its own body in its own camera through published retinal biology running
on hardware. The reafference ladder M0 to M3 is the engineering sequence: discretize action with
a move-stop policy, let the body report its own arrival through proprioceptive timing, make arrival
predictable, and identify the body as a place through sensitize-adapt dynamics. The retinal circuit
is the one measured in salamander retina by Kastner and Baccus (2013) and others, with thirteen
published constants carried at their measured values. Nothing in the chain is tuned. The release
time constant arrives on our camera at 2.35 seconds against the paper’s 2.40, over 98 continuous
movement offsets in a novel room with direct sun and moving shadow. The body falls out as the
region of the image that stays still while everything touching it moves.
No markers, no CAD
model, no training set. The talk will present the implementation, the pre-registered falsification
conditions, and the honest limits. 𝜏_on fails against the instrument ceiling.
One parameter is
not derived. The core observation worth the room’s attention is that the policy chosen at M0 for
credit assignment also solves self-localization at M3. One policy. Two problems usually treated
as separate.
Active inference predicts this in principle through information-maximizing policy
selection, but the specific case of move-stop resolving both at once appears to be new to the
literature. Pre-recorded with embedded demonstration video. Live Q&A after broadcast if the
symposium runs one.
59

## Page 60

55
ROIF and Active Inference:
The Formal Boundary of
History-Conditioned Anticipatory Systems.
History-
dependent
reconfiguration,
structured
memory,
and
predictive preconfiguration in prestressed complex sys-
tems
Presenter: Vitalii Shapovalov
Contribution: Pre-recorded Video presentation
Abstract
Recursive Organic Integration Framework (ROIF) is a deterministic, domain-agnostic computa-
tional architecture for complex systems whose present organization and subsequent responses
may be shaped by prior interactions, distributed reconfiguration, and anticipatory preconfigura-
tion.
Rather than treating history as an additional input variable, ROIF represents how prior
events may become embodied in the organization through which subsequent events are processed.
The framework distinguishes experimentally separable routes through which the past and repre-
sented possible futures may affect subsequent dynamics: history embodied in present organiza-
tion; structured memory conditioning a subsequent transition; and anticipatory preconfiguration
based on a represented possible future Temporal Image. Matched-current-state controls found no
independently identifiable history-specific transition effect for the tested physical-history pathway.
In contrast, a separate structured-memory experiment produced different subsequent physical
transitions under matched represented physical state and adaptive connection state. The predic-
tive architecture allows possible future SystemImages to participate in present preconfiguration.
Current experiments, however, test only a restricted one-step preconfiguration within the imple-
mented prestress representation. This experimental restriction does not define the broader archi-
tectural role of prestress in ROIF, where prestress reflects an already organized relational system
state shaped by interacting internal and external conditions, constraints, capacities, history, and
time. Objective-independence tests further prevent interpreting the present result as evidence for
general predictive stabilization. These results motivate a boundary question rather than a claim
of equivalence: when does a history-conditioned anticipatory architecture constitute Active Infer-
ence? ROIF does not presently require a variational free-energy functional, Bayesian generative
model, or particular neurobiological implementation. The presentation therefore uses ROIF as a
computational case study to ask which mathematical conditions distinguish Active Inference from
a broader class of history-conditioned anticipatory systems, and whether these formalisms should
be understood as nested, neighbouring, or formally distinct.
60

## Page 61

56
Modelling Language Comprehension with Embodied Ac-
tive Inference
Presenter: Vitória de Souza
Contribution: Pre-recorded Video presentation
Abstract
How do we understand sentences like “The boy is opening a can of soda”? Accounts of grounded
and embodied cognition propose that comprehension is supported by the partial reactivation of
sensorimotor states. For instance, understanding the “The boy is opening a can of soda” entails re-
activating motor and visual perceptual states congruent with the action described (i.e., pinching).
While verbal theories have been offered to account for sensorimotor contributions to comprehen-
sion, mathematical and mechanistic models are scant. To address this gap, this study provides
a proof-of-concept, hierarchical active inference model of sensorimotor contributions to sentence
comprehension. In our study, we simulate sensorimotor simulations in a seminal sentence-picture
verification task in which participants read a sentence and then are shown a picture of an agent
performing an action that either matches or mismatches the shape of the action implied by the
sentence. Behaviourally, participants are slower at deciding that the picture reflects the concepts
of the sentence when the shape mismatches, implicating sensorimotor reactivation. Our model
simulations show that mismatching shapes cause extended epistemic foraging, leading to slower
reaction times. This model provides a mathematical and mechanistic account of embodied cogni-
tion, extending verbal theories.
61

## Page 62

57
From Lived Experience to Mathematical Mapping: A Com-
putational Phenomenology of Trauma Recovery
Presenter: Xin Sun
Contribution: Pre-recorded Video presentation
Abstract
Severe childhood maltreatment is known to impair brain development. However, this retrospective
single-case autoethnography illustrates that an individual with extreme developmental trauma can
achieve functional recovery without formal intervention. This study covers the participant’s life
from birth to age 28, and the core recovery period begins with cognitive collapse at age 14 and ends
with high functional recovery at age 28. This recovery is marked by the participant‑researcher’s
ability to produce this manuscript. Consequently, this case shapes a distinct recovery trajectory
that existing models do not capture.
62