pith. sign in

arxiv: 2606.22813 · v1 · pith:WHAOC5LTnew · submitted 2026-06-22 · 💻 cs.AI

Active Inference as the Test-Time Scaling Law for Physical AI Agents

Omar Hashash , Christo Kurisummoottil Thomas , Walid Saad , Merouane Debbah , Karl Friston , Adeel Razi This is my paper

Pith reviewed 2026-06-26 08:52 UTC · model grok-4.3

classification 💻 cs.AI
keywords active inferencetest-time scalingphysical AI agentsBayesian policy updatevariational free energygeneralizationautonomous drivingworld model
0
0 comments X

The pith

Active inference supplies a test-time scaling law for physical AI agents by updating policies through soft Bayesian inference on prediction errors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that active inference, based on the principle of surviving in the real world, supplies the reasoning agents need to resolve prediction errors when they encounter situations outside their training distribution. The derived scaling law models policy updates at test time as a soft Bayesian inference process in which allowable policies that reduce expected errors serve as the likelihood, producing a posterior policy. A variational solution that minimizes free energy bounds makes the update tractable and extends learning by reinforcing newly resolved instances in both the policy and the world model. The approach therefore scales with an agent's ongoing real-world experience rather than model size or training data volume. A sympathetic reader would care because it offers a route to robust generalization for physical agents such as autonomous vehicles in non-stationary settings where conventional scaling methods reach their limits.

Core claim

The paper states that the first principle of active inference equips physical AI agents with the general objective to survive, under which specific task objectives are subsumed, and that this principle supplies the reasoning to resolve prediction errors outside the training distribution. The scaling law captures the process by dynamically updating the agent's policy at test time through a soft Bayesian inference step whose likelihood is the reasoning that reduces expected prediction errors under allowable policies. The resulting posterior policy recovers a biological scaling mechanism that engages the basal ganglia and prefrontal cortex. An analytically tractable variational inference soluti

What carries the argument

Soft Bayesian inference process for policy update, in which beliefs about the policy are revised using reasoning that reduces expected prediction errors under allowable policies as the likelihood, solved via variational free-energy minimization.

If this is right

  • Policy updates at test time enable generalization in non-stationary environments.
  • The variational solution reinforces new instances in both the policy and the world model.
  • The posterior policy recovers the scaling mechanism engaging the basal ganglia and prefrontal cortex.
  • The method scales with continuous real-world experience rather than model size or training data.
  • Simulation results show robust generalization and over 36 percent improvement in inference efficiency on autonomous driving tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same variational update rule could be applied to other embodied tasks such as robotic manipulation without requiring task-specific retraining.
  • The biological mapping suggests that hardware implementations might benefit from architectures that separate policy evaluation from world-model updating in the manner of basal ganglia and prefrontal circuits.
  • Because the scaling is driven by real-time experience, the law predicts continued performance gains for agents that remain deployed for long periods rather than for agents that are periodically retrained offline.

Load-bearing premise

Active inference equips agents with the general objective to survive in the real world under which specific task objectives are subsumed, and the variational inference solution extends to enable learning beyond training by reinforcing new instances resolved at test time in both the policy and world model.

What would settle it

A controlled autonomous-driving simulation in which an agent equipped with the variational free-energy policy update shows no improvement in generalization to unforeseen scenarios or no gain in inference efficiency relative to model-free Q-learning or model-based Bayesian reinforcement learning would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.22813 by Adeel Razi, Christo Kurisummoottil Thomas, Karl Friston, Merouane Debbah, Omar Hashash, Walid Saad.

Figure 1
Figure 1. Figure 1: Illustration of the solutions for survival at test time by conserving the NESS of physical AI agents, according to the different underlying conditions assumed about the world. of “biological” reasoning at test time in physical AI agents, to date, might possibly explain their failures and limited capacity to survive in the real world [9]. Indeed, the physical world is full of unforeseen scenarios that an AI… view at source ↗
Figure 2
Figure 2. Figure 2: Test-time scaling law for physical AI agents (top) vs. test-time scaling law in LLMs (bottom). The proposed framework grounds test-time scaling in active inference, where prediction errors trigger the transition from BG execution to PFC reasoning, followed by a feedback to scale the policy in the BG. Here, the LLM’s (e.g., GPT-5) routing and scaling represent a special empirical case of the proposed framew… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of our test-time scaling framework comprising a physical AI agent (e.g., an autonomous vehicle) that reasons through the world model at the wireless network edge to generalize in unforeseen scenarios. As the world evolves according to the causal structure of the environment, the states of the elements at time τ are causally dependent on their states in the previous time instant τ − 1 and the c… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the test-time scaling mechanism in which the AI agent must generalize in an unforeseen scenario that considers a jaywalking pedestrian appearing at test time. from the BG which governs policy execution to the PFC that initiates deliberative reasoning. In particular, the PFC responds by transitioning to active inference which includes planning (as inference) that generates counterfactual sim… view at source ↗
Figure 5
Figure 5. Figure 5: The roadmap for the proposed variational inference solution that unifies perception, planning, action, and learning to enable the test-time scaling framework for the physical AI agent. in terms of free energy minimization that essentially serves as a bound on (log) model evidence or surprisal. Intuitively, action and learning should further follow this free energy formulation as they can be modeled in term… view at source ↗
Figure 6
Figure 6. Figure 6: Forney-style factor graph representing the generative model of the POMDP. as an upper bound on prediction errors. We thus define a POMDP to model this problem, with the following key components: • Agent: The network is the agent that performs inference and reasoning on behalf of the physical AI agent in the real world. In other words, the AI agent offloads its reasoning process onto the network edge. • Obs… view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of the Markov Blanket that statistically separates the AI agent from its external environment. that views the AI agent as a random dynamical system in interaction with its environment. Using state flow methods from statistical physics, we can then analyze the behavior of this system as it maintains NESS to survive. In particular, this flow can describe how the policy πo must update based on ψ(… view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of the gradient descent operation over the Lagrangian to capture the scaled policy as derived in (20) and (21). prediction error as ϵπo,ψ = ∇aF(s, a, o) = ln ψ(a | st ,π) − ln π ′ o (a | st). Here, the posterior beliefs about the actions are obtained through a gradient descent as in (13). Starting initially from π ′ o (a | st) = πo(a | st), the scaled policy π ′ (a | st) can then be reached th… view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of the learning process that is modeled as inference to reinforce the new experiences discovered at test time to enable continually learning the policy of the physical AI agent and the world model. distribution with concentration parameters ν ′ ·j = [ν ′ 1j , . . . , ν′ |A|j ] over each column j, which factorizes into: q ∗ (πo) = Y |S| j=1 Dir(Π˜ ·j | ν ′ ·j ). (41) This showcases how updating… view at source ↗
Figure 10
Figure 10. Figure 10: Comparison between (a) Q-learning and (b) the proposed test-time scaling on an unforeseen jaywalking scenario. In (a), the Q-learning agent crashes at Step 2 upon the pedestrian’s appearance, failing to adapt its policy. In (b), the proposed scaling law detects the pedestrian as an unforeseen scenario at Step 2, triggering a switch from feed-forward (blue) to scaled (yellow) actions, allowing the agent to… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison between the rewards of (a) Q-learning and (b) the proposed test-time scaling over the time steps in the unforeseen scenario, with the variation of surprise levels during policy scaling in this scenario. during training23. However, as the pedestrian appears at τ = 2 having L = 1, surprise abruptly increases to reach θ = 0.29 and crosses the threshold ϵ. This signals an unforeseen scenario has oc… view at source ↗
Figure 12
Figure 12. Figure 12: Performance comparison in terms of training rewards, test-time rewards (in unforeseen scenario), success rate, and inference rate between Q-learning, Bayesian RL, and the proposed test-time scaling method. accelerates to reach its goal at τ = 14 with cumulative reward r = 22.9. Clearly, this shows how our proposed test-time scaling method generalizes to unforeseen scenarios while outperforming Q-learning … view at source ↗
Figure 13
Figure 13. Figure 13: (a) The change in surprise and update of the (b) world model and (c) policy parameters through inference while encountering the resolved unforeseen scenarios over 50 episodes. update such that the unforeseen scenario becomes progressively less surprising over episodes. Asymptotically, the surprise S approaches S ≈ 0.2 bits by episode 50. In this case, the agent confidently expects a pedestrian to appear i… view at source ↗
read the original abstract

In this paper, a novel test-time scaling law for physical artificial intelligence (AI) agents is introduced. This scaling law enables physical AI agents to reason with their world models to generalize in unforeseen scenarios at test time. The derived scaling law is grounded in the first principle of active inference, which equips agents with the general objective to survive in the real world, under which their specific task objectives are subsumed. Active inference achieves this by providing the reasoning to resolve prediction errors that arise when the agent encounters unforeseen situations outside its training distribution, enabling generalization in non-stationary environments. The proposed scaling law captures this by dynamically updating the agent's policy with this reasoning at test time. This policy update is modeled as a soft Bayesian inference process in which beliefs about the policy are updated using the reasoning that reduces expected prediction errors under allowable policies as a likelihood. The resulting posterior policy admits a biological interpretation, recovering the scaling mechanism that engages the brain's basal ganglia and prefrontal cortex at test time. To solve this analytically intractable problem, a variational inference solution minimizing free energy bounds is developed. This solution extends to enable learning beyond training by reinforcing new instances, resolved at test time, in both the policy and world model. Unlike existing scaling laws constrained by model size and training data, the derived solution scales with the continuous real-world experience of a physical AI agent. Simulation results on an autonomous driving task demonstrate that the proposed solution outperforms model-free Q-learning and model-based Bayesian reinforcement learning, achieving robust generalization to unforeseen scenarios while improving inference efficiency by over 36%.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims to derive a test-time scaling law for physical AI agents from active inference as a first principle. Agents use variational free-energy minimization to perform soft Bayesian policy updates at test time, resolving prediction errors for generalization in non-stationary settings. The solution is asserted to extend beyond inference to reinforce resolved instances in both the policy and world model, enabling scaling with continuous experience rather than training data. A biological interpretation links the posterior policy to basal ganglia and prefrontal cortex mechanisms. Simulations on an autonomous driving task report outperformance over model-free Q-learning and model-based Bayesian RL, plus >36% inference efficiency gains.

Significance. If the variational extension to test-time world-model reinforcement holds with explicit, stable mechanisms, the work would supply a principled alternative to data- or parameter-size scaling laws for embodied agents, with direct relevance to robust physical AI in open environments. The grounding in active inference supplies a unified objective (survival under prediction-error minimization) that subsumes task-specific goals, and the simulation comparison provides an initial empirical anchor.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'this solution extends to enable learning beyond training by reinforcing new instances, resolved at test time, in both the policy and world model' is load-bearing for the distinction from training-constrained scaling laws, yet no explicit update rule, gradient, or outer optimization loop for world-model parameters is supplied. Variational free-energy minimization is conventionally an inference procedure over latents or policies; parameter learning requires additional structure (e.g., EM-style alternation or online gradients on the bound) whose stability under non-stationary test-time data is not addressed.
  2. [Abstract] Abstract (simulation paragraph): The reported outperformance and 'improving inference efficiency by over 36%' are presented without error bars, number of independent runs, statistical tests, or precise definition of the efficiency metric. Because the scaling-law claim rests on these results demonstrating generalization beyond training, the absence of these details prevents assessment of whether gains arise from the proposed mechanism or from other implementation choices.
minor comments (2)
  1. [Abstract] The abstract states that an 'analytically intractable problem' is solved variationally but supplies neither the intractable objective nor the free-energy bound that is minimized, making it impossible for readers to verify the derivation steps.
  2. [Abstract] Notation for the soft Bayesian policy update (beliefs over policies updated by 'reasoning that reduces expected prediction errors under allowable policies as a likelihood') is introduced without an equation; a compact mathematical statement would clarify the mapping from active-inference quantities to the claimed posterior.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract claims and empirical reporting. We address each major comment below and will incorporate clarifications and additions in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'this solution extends to enable learning beyond training by reinforcing new instances, resolved at test time, in both the policy and world model' is load-bearing for the distinction from training-constrained scaling laws, yet no explicit update rule, gradient, or outer optimization loop for world-model parameters is supplied. Variational free-energy minimization is conventionally an inference procedure over latents or policies; parameter learning requires additional structure (e.g., EM-style alternation or online gradients on the bound) whose stability under non-stationary test-time data is not addressed.

    Authors: We agree that the extension to world-model reinforcement requires explicit formalization to support the test-time scaling claim. The manuscript centers on variational free-energy minimization for policy inference, with the world-model extension stated at a high level in the abstract. In revision we will add a new subsection deriving the online gradient update on the variational bound for world-model parameters (following an EM-style alternation between policy and model optimization) and include a brief analysis of stability under non-stationary test-time data streams. This will make the mechanism concrete without altering the core active-inference derivation. revision: yes

  2. Referee: [Abstract] Abstract (simulation paragraph): The reported outperformance and 'improving inference efficiency by over 36%' are presented without error bars, number of independent runs, statistical tests, or precise definition of the efficiency metric. Because the scaling-law claim rests on these results demonstrating generalization beyond training, the absence of these details prevents assessment of whether gains arise from the proposed mechanism or from other implementation choices.

    Authors: We concur that the simulation results must be reported with full statistical detail to substantiate the generalization claims. The current version reports aggregate metrics only. In the revision we will (i) specify the inference-efficiency metric as average wall-clock time per policy update, (ii) report means and standard errors over 10 independent random seeds, and (iii) include paired t-tests against the Q-learning and Bayesian-RL baselines. These additions will be placed in both the abstract and the results section. revision: yes

Circularity Check

2 steps flagged

Active inference scaling law reduces to authors' own free-energy framework by construction via self-citation and asserted extension

specific steps
  1. self citation load bearing [Abstract]
    "The derived scaling law is grounded in the first principle of active inference, which equips agents with the general objective to survive in the real world, under which their specific task objectives are subsumed. ... The resulting posterior policy admits a biological interpretation, recovering the scaling mechanism that engages the brain's basal ganglia and prefrontal cortex at test time."

    Active inference and its free-energy principle originate in prior work by co-author Karl Friston; the paper invokes this as external first principle and recovers the biological mechanism from the same literature, so the scaling law's grounding reduces to the authors' own framework rather than independent justification.

  2. self definitional [Abstract]
    "To solve this analytically intractable problem, a variational inference solution minimizing free energy bounds is developed. This solution extends to enable learning beyond training by reinforcing new instances, resolved at test time, in both the policy and world model. Unlike existing scaling laws constrained by model size and training data, the derived solution scales with the continuous real-world experience of a physical AI agent."

    The scaling law is defined as the dynamic policy update via free-energy minimization; the asserted extension to test-time world-model reinforcement is presented as following from the same variational solution, rendering the 'new' scaling behavior equivalent to the input active-inference objective by construction.

full rationale

The paper's central derivation grounds the test-time scaling law directly in active inference as a 'first principle' (with co-author Friston as originator), models policy update as soft Bayesian inference under free-energy minimization, and asserts without separate mechanism that the variational solution 'extends to enable learning beyond training by reinforcing new instances... in both the policy and world model.' This makes the claimed departure from training-data scaling laws equivalent to re-applying the input framework at test time rather than deriving an independent result. One load-bearing self-citation chain and self-definitional extension are present; the simulation outperformance does not rescue the logical reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on active inference as the foundational principle and the validity of the variational free-energy minimization as a solution to the intractable Bayesian update; no explicit free parameters or new entities are introduced in the abstract.

axioms (2)
  • domain assumption Active inference provides the general objective to survive in the real world under which specific task objectives are subsumed.
    Invoked in the abstract as the grounding for the scaling law and generalization mechanism.
  • domain assumption The variational inference solution minimizing free energy bounds solves the analytically intractable soft Bayesian policy update.
    Stated as the developed solution that extends to test-time learning.

pith-pipeline@v0.9.1-grok · 5830 in / 1510 out tokens · 17579 ms · 2026-06-26T08:52:29.554575+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 7 linked inside Pith

  1. [1]

    Artificial general intelligence (AGI)-native wireless systems: A journey beyond 6G,

    W. Saad, O. Hashash, C. K. Thomas, C. Chaccour, M. Debbah, N. Mandayam, and Z. Han, “Artificial general intelligence (AGI)-native wireless systems: A journey beyond 6G,”Proceedings of the IEEE, vol. 113, no. 9, pp. 849–887, 2025

    2025

  2. [2]

    Waymos blocked roads and caused chaos during San Francisco power outage,

    J. Ding and M. Liedtke, “Waymos blocked roads and caused chaos during San Francisco power outage,” https://fortune.com/ 2025/12/22/waymo-ai-san-francisco-power-outage-operational-management-failure-software/, December 2025, associated Press, December 22, 2025

    2025

  3. [3]

    World models,

    D. Ha and J. Schmidhuber, “World models,”arXiv preprint arXiv:1803.10122, vol. 2, no. 3, p. 440, 2018

    Pith/arXiv arXiv 2018

  4. [4]

    Kahneman,Thinking, fast and slow

    D. Kahneman,Thinking, fast and slow. macmillan, 2011

    2011

  5. [5]

    A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27,

    Y . LeCun, “A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27,”Open Review, vol. 62, 2022

    2022

  6. [6]

    Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects,

    R. P. Rao and D. H. Ballard, “Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects,”Nature neuroscience, vol. 2, no. 1, pp. 79–87, 1999

    1999

  7. [7]

    Active inference: a process theory,

    K. Friston, T. FitzGerald, F. Rigoli, P. Schwartenbeck, and G. Pezzulo, “Active inference: a process theory,”Neural computation, vol. 29, no. 1, pp. 1–49, 2017

    2017

  8. [8]

    Dissipative structures in biological systems: bistability, oscillations, spatial patterns and waves,

    A. Goldbeter, “Dissipative structures in biological systems: bistability, oscillations, spatial patterns and waves,”Philosoph- ical transactions. Series A, Mathematical, physical, and engineering sciences, vol. 376, no. 2124, p. 20170376, 2018

    2018

  9. [9]

    Active inference for physical AI agents–an engineering perspective,

    B. de Vries, “Active inference for physical AI agents–an engineering perspective,”arXiv preprint arXiv:2603.20927, 2026

    arXiv 2026

  10. [10]

    Life as we know it,

    K. Friston, “Life as we know it,”Journal of the royal society interface, vol. 10, no. 86, p. 20130475, 2013. 52

    2013

  11. [11]

    Challenges of real-world reinforcement learning: definitions, benchmarks and analysis,

    G. Dulac-Arnold, N. Levine, D. J. Mankowitz, J. Li, C. Paduraru, S. Gowal, and T. Hester, “Challenges of real-world reinforcement learning: definitions, benchmarks and analysis,”Machine Learning, vol. 110, no. 9, pp. 2419–2468, 2021

    2021

  12. [12]

    OpenAI, “GPT-5,” https://openai.com/, 2025

    2025

  13. [13]

    V- JEPA 2: Self-supervised video models enable understanding, prediction and planning,

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholuset al., “V- JEPA 2: Self-supervised video models enable understanding, prediction and planning,”arXiv preprint arXiv:2506.09985, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Mindjourney: Test-time scaling with world models for spatial reasoning,

    Y . Yang, J. Liu, Z. Zhang, S. Zhou, R. Tan, J. Yang, Y . Du, and C. Gan, “Mindjourney: Test-time scaling with world models for spatial reasoning,”Advances in Neural Information Processing Systems, vol. 38, pp. 109 855–109 885, 2026

    2026

  15. [15]

    π 0.7: a steerable generalist robotic foundation model with emergent capabilities,

    Physical Intelligence, “π 0.7: a steerable generalist robotic foundation model with emergent capabilities,” Physical Intelligence, Tech. Rep., 2026. [Online]. Available: https://pi.website/pi07

    2026

  16. [16]

    V-JEPA 2.1: Unlocking dense features in video self-supervised learning,

    L. Mur-Labadia, M. Muckley, A. Bar, M. Assran, K. Sinha, M. Rabbat, Y . LeCun, N. Ballas, and A. Bardes, “V-JEPA 2.1: Unlocking dense features in video self-supervised learning,”arXiv preprint arXiv:2603.14482, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    The markov blankets of life: autonomy, active inference and the free energy principle,

    M. Kirchhoff, T. Parr, E. Palacios, K. Friston, and J. Kiverstein, “The markov blankets of life: autonomy, active inference and the free energy principle,”Journal of The royal society interface, vol. 15, no. 138, p. 20170792, 2018

    2018

  18. [18]

    A neural substrate of prediction and reward,

    W. Schultz, P. Dayan, and P. R. Montague, “A neural substrate of prediction and reward,”Science, vol. 275, no. 5306, pp. 1593–1599, 1997

    1997

  19. [19]

    Training compute-optimal large language models,

    J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherford, D. Casas, L. A. Hendricks, J. Welbl, A. Clark et al., “Training compute-optimal large language models,”arXiv preprint arXiv:2203.15556, vol. 10, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,”arXiv preprint arXiv:2001.08361, 2020

    Pith/arXiv arXiv 2001

  21. [21]

    Scaling LLM test-time compute optimally can be more effective than scaling model parameters,

    C. Snell, J. Lee, K. Xu, and A. Kumar, “Scaling LLM test-time compute optimally can be more effective than scaling model parameters,”arXiv preprint arXiv:2408.03314, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Toward causal representation learning,

    B. Sch ¨olkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y . Bengio, “Toward causal representation learning,”Proceedings of the IEEE, vol. 109, no. 5, pp. 612–634, 2021

    2021

  23. [23]

    From passive mirrors to active agents: Holonic digital twins for physical artificial intelligence over networks,

    C. K. Thomas, O. Hashash, and W. Saad, “From passive mirrors to active agents: Holonic digital twins for physical artificial intelligence over networks,”IEEE Vehicular Technology Magazine, 2026

    2026

  24. [24]

    Robust agents learn causal world models,

    J. Richens and T. Everitt, “Robust agents learn causal world models,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  25. [25]

    Bayesian surprise attracts human attention,

    L. Itti and P. Baldi, “Bayesian surprise attracts human attention,”Vision research, vol. 49, no. 10, pp. 1295–1306, 2009

    2009

  26. [26]

    The free-energy principle: a unified brain theory?

    K. Friston, “The free-energy principle: a unified brain theory?”Nature reviews neuroscience, vol. 11, no. 2, pp. 127–138, 2010

    2010

  27. [27]

    T. Parr, G. Pezzulo, and K. J. Friston,Active inference: the free energy principle in mind, brain, and behavior. MIT Press, 2022

    2022

  28. [28]

    Path integrals, particular kinds, and strange things,

    K. Friston, L. Da Costa, D. A. Sakthivadivel, C. Heins, G. A. Pavliotis, M. Ramstead, and T. Parr, “Path integrals, particular kinds, and strange things,”Physics of Life Reviews, vol. 47, pp. 35–62, 2023

    2023

  29. [29]

    Evidence for surprise minimization over value maximization in choice behavior,

    P. Schwartenbeck, T. H. FitzGerald, C. Mathys, R. Dolan, M. Kronbichler, and K. Friston, “Evidence for surprise minimization over value maximization in choice behavior,”Scientific reports, vol. 5, no. 1, p. 16575, 2015

    2015

  30. [30]

    Pearl,Probabilistic reasoning in intelligent systems: networks of plausible inference

    J. Pearl,Probabilistic reasoning in intelligent systems: networks of plausible inference. Elsevier, 2014

    2014

  31. [31]

    The mathematics of changing one’s mind, via jeffrey’s or via pearl’s update rule,

    B. Jacobs, “The mathematics of changing one’s mind, via jeffrey’s or via pearl’s update rule,”Journal of Artificial Intelligence Research, vol. 65, pp. 783–806, 2019

    2019

  32. [32]

    R. S. Sutton and A. G. Barto,Reinforcement Learning: An Introduction. Cambridge, MA: MIT Press, 2018. 53

    2018

  33. [33]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovskiet al., “Human-level control through deep reinforcement learning,”nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  34. [34]

    Optimizing for the future in non- stationary mdps,

    Y . Chandak, G. Theocharous, S. Shankar, M. White, S. Mahadevan, and P. Thomas, “Optimizing for the future in non- stationary mdps,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 1414–1425

    2020

  35. [35]

    A tutorial on energy-based learning,

    Y . LeCun, S. Chopra, R. Hadsell, M. Ranzato, F. Huanget al., “A tutorial on energy-based learning,”Predicting structured data, vol. 1, no. 0, 2006

    2006

  36. [36]

    An introduction to variational methods for graphical models,

    M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul, “An introduction to variational methods for graphical models,” Machine learning, vol. 37, no. 2, pp. 183–233, 1999

    1999

  37. [37]

    Graphical models, exponential families, and variational inference,

    M. J. Wainwright and M. I. Jordan, “Graphical models, exponential families, and variational inference,”Foundations and Trends® in Machine Learning, vol. 1, no. 1-2, pp. 1–305, 2008

    2008

  38. [38]

    Variational message passing

    J. Winn, C. M. Bishop, and T. Jaakkola, “Variational message passing.”Journal of Machine Learning Research, vol. 6, no. 4, 2005

    2005

  39. [39]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  40. [40]

    A step-by-step tutorial on active inference and its application to empirical data,

    R. Smith, K. J. Friston, and C. J. Whyte, “A step-by-step tutorial on active inference and its application to empirical data,” Journal of mathematical psychology, vol. 107, p. 102632, 2022

    2022

  41. [41]

    Codes on graphs: Normal realizations,

    G. D. Forney, “Codes on graphs: Normal realizations,”IEEE Transactions on Information Theory, vol. 47, no. 2, pp. 520–548, 2001

    2001

  42. [42]

    C. M. Bishop and N. M. Nasrabadi,Pattern recognition and machine learning. Springer, 2006, vol. 4, no. 4

    2006

  43. [43]

    Whence the expected free energy?

    B. Millidge, A. Tschantz, and C. L. Buckley, “Whence the expected free energy?”Neural Computation, vol. 33, no. 2, pp. 447–482, 2021

    2021

  44. [44]

    Koller and N

    D. Koller and N. Friedman,Probabilistic graphical models: principles and techniques. MIT press, 2009

    2009

  45. [45]

    Bayesian mechanics for stationary processes,

    L. Da Costa, K. Friston, C. Heins, and G. A. Pavliotis, “Bayesian mechanics for stationary processes,”Proceedings. Mathematical, Physical, and Engineering Sciences, vol. 477, no. 2256, p. 20210518, 2021

    2021

  46. [46]

    On markov blankets and hierarchical self-organisation,

    E. R. Palacios, A. Razi, T. Parr, M. Kirchhoff, and K. Friston, “On markov blankets and hierarchical self-organisation,” Journal of theoretical biology, vol. 486, p. 110089, 2020

    2020

  47. [47]

    Nonlinear kinetics on lattices based on the kinetic interaction principle,

    G. Kaniadakis and D. T. Hristopulos, “Nonlinear kinetics on lattices based on the kinetic interaction principle,”Entropy, vol. 20, no. 6, p. 426, 2018

    2018

  48. [48]

    Cognitive dynamics: From attractors to active inference,

    K. Friston, B. Sengupta, and G. Auletta, “Cognitive dynamics: From attractors to active inference,”Proceedings of the IEEE, vol. 102, no. 4, pp. 427–445, 2014

    2014

  49. [49]

    Welcome to the era of experience,

    D. Silver and R. S. Sutton, “Welcome to the era of experience,”Google AI, vol. 1, p. 11, 2025

    2025

  50. [50]

    Active inference on discrete state-spaces: A synthesis,

    L. Da Costa, T. Parr, N. Sajid, S. Veselic, V . Neacsu, and K. Friston, “Active inference on discrete state-spaces: A synthesis,” Journal of Mathematical Psychology, vol. 99, p. 102447, 2020

    2020

  51. [51]

    Why AI systems don’t learn and what to do about it: Lessons on autonomous learning from cognitive science,

    E. Dupoux, Y . LeCun, and J. Malik, “Why AI systems don’t learn and what to do about it: Lessons on autonomous learning from cognitive science,”arXiv preprint arXiv:2603.15381, 2026

    arXiv 2026

  52. [52]

    Edge continual learning for dynamic digital twins over wireless networks,

    O. Hashash, C. Chaccour, and W. Saad, “Edge continual learning for dynamic digital twins over wireless networks,” in Proc. of the IEEE 23rd International Workshop on Signal Processing Advances in Wireless Communication (SPAWC), Oulu, Finland, Jul. 2022, pp. 1–5

    2022

  53. [53]

    The neural basis of the speed–accuracy tradeoff,

    R. Bogacz, E.-J. Wagenmakers, B. U. Forstmann, and S. Nieuwenhuis, “The neural basis of the speed–accuracy tradeoff,” Trends in neurosciences, vol. 33, no. 1, pp. 10–16, 2010

    2010