Active Inference for Self-Organizing Multi-LLM Systems: A Bayesian Thermodynamic Approach to Adaptation

Rithvik Prakki

REVIEW 3 major objections 4 minor 1 cited by

Active Inference for Self-Organizing Multi-LLM Systems: A Bayesian Thermodynamic Approach to Adaptation

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read An active inference layer can learn which prompts produce good LLM output.

desk verdict A genuinely new application of active inference to prompt and search selection for LLM agents, with released code, but the evidence is qualitative and the expected-free-energy derivation contains a false identity, so the central claims are not yet supported. read the letter →

arxiv 2412.10425 v3 pith:XFJDEBFS submitted 2024-12-10 cs.CL cs.AI

Rithvik Prakki This is my paper · ORCID

classification cs.CLcs.AI

keywords activeinferencefreeenergyprinciplelargelanguagemodelspromptadaptationexploration-exploitationBayesianlearningmulti-agentsystemsexpected

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that active inference, a Bayesian scheme that minimizes variational free energy by updating beliefs and choosing actions, can act as an adaptive layer over a multi-LLM research agent. The agent models its environment with three hidden state factors—prompt, search, and information state—and receives seven quality-scored observation modalities, then learns which prompt combinations and search actions produce strong outputs. The authors report that the learned observation matrices develop clear structure and that action selection shifts from search-heavy exploration to prompt-focused exploitation over time. If this holds, LLM-based agents could adapt their own prompts and strategies without retraining, guided by a principled information-seeking objective.

What carries the argument

The engine of the system is the active inference generative model: observation matrices A that map hidden prompt, search, and information states to seven quality modalities, transition matrices B that govern state changes under actions, preference matrix C encoding which observations are desirable, and Dirichlet-based parameter learning that updates A and B after each observation. Policy selection minimizes expected free energy Gπ, decomposed into information gain and pragmatic value, and applies a softmax over policies. This machinery is what converts raw LLM quality scores into beliefs about which prompts work and what to try next.

What would settle it

Run the same active inference loop with two different evaluator LLMs (or with human judges) on the same prompt and search space; if the learned observation matrices and the ranking of preferred prompts diverge substantially across evaluators, the claim that the agent has learned accurate environment dynamics rather than evaluator idiosyncrasies is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that an active inference agent placed above an LLM-based research agent can learn an accurate generative model of the prompt/search environment from structured evaluation scores alone. The evidence offered is the emergence of non-uniform structure in the learned observation matrices, which now distinguish high-scoring prompt and search states from poor ones, and a temporal action pattern in which early search-dominant policies give way to targeted prompt testing. The paper presents this as a demonstration that expected free energy minimization generalizes active inference from low-dimensional control problems to high-dimensional, language-driven agent adaptation.

Load-bearing premise

The central claim collapses if the LLM evaluator's structured quality scores are not reliable, consistent observations of a fixed environment, because the agent learns only from those scores and there is no external ground truth in the experiments.

Editorial extensions

If this is right

LLM agents can adapt their own prompts and search strategies during deployment without gradient updates or fine-tuning.
The early exploration-to-exploitation shift emerges from the free energy objective, not from a hand-coded schedule.
The learned observation matrices provide an inspectable record of which prompt and search states the agent believes produce high-quality outputs.
The framework requires only structured, machine-readable quality feedback, making it compatible with LLM evaluators that output JSON scores.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the 'environment' is defined by the GPT-4o-mini evaluator, the learned structure may reflect the evaluator's scoring tendencies; a fair test would compare learned prompt rankings against human or held-out ground truth.
A neighbouring question the paper leaves open is whether EFE-guided prompt selection beats simpler baselines such as random search or upper-confidence-bound bandits; the reported experiments do not include that comparison.
Scaling to larger prompt spaces would likely require factored transition tensors or hierarchical models, since the current 33×33×33 prompt transition tensor grows cubically with the number of prompt states.
The exploration-first pattern may be shaped by the strong preference prior on detailed information states; adjusting that prior could shift the exploration-exploitation balance and is a testable extension.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is applying active inference to select prompts and search actions for a multi-LLM research agent, and the author does ship code and a fairly detailed generative-model setup. That is real, and it distinguishes this from the usual LLM-agent papers. The related work is honest and the state-factor decomposition (prompt, search, information) is a reasonable first cut. I also think the idea of using an active-inference layer instead of yet another RL loop is worth taking seriously.

The soft spots are real, though. The stress-test note is right: Eq. 10 in Section 3.3 is not Bayes' theorem. The correct identity is q(s|o,pi)q(o|pi) = q(o|s,pi)q(s|pi), and what the paper writes instead does not follow. Since the expected-free-energy objective in Section 4.5 is the stated basis for the claimed exploration-exploitation behavior, this is a load-bearing error, not a typo. The author cites standard active-inference sources and the formula itself is familiar, so a sympathetic reader could repair the derivation, but as written the theoretical justification does not hold.

The empirical section is the bigger gap. There are no baselines, no quantitative metrics, no error bars, and no external ground truth. The "environment" is the paper's own GPT-4o-mini evaluator, so the claim that the agent learned "accurate models of environment dynamics" is largely self-referential: the learned matrices may just mirror the evaluator's scoring tendencies. Figures 2-5 show structure, but visual inspection alone does not establish that the agent learned anything useful, nor that the action pattern is driven by free-energy minimization rather than by the hand-chosen preference matrix and learning rates. The thermodynamic framing in the introduction is also mostly rhetorical; it does not do quantitative work in the experiments.

Who is this for? Researchers working on adaptive LLM agents and maybe the active-inference community. It is an exploratory paper with a new application and reproducible code, but the current evidence would not justify accepting it as a strong archival result. It deserves a serious referee, though: the idea is novel enough and the code is available, and external feedback could turn this into something useful. I would send it to review rather than desk-reject, with the expectation that the author must add baselines, quantify the results, and fix the derivation. My own verdict is skeptical, but this is a paper to engage with, not dismiss.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes an active inference layer as a cognitive controller for a multi-LLM research agent. The agent maintains discrete state factors for prompt, search, and information states, receives seven observation modalities (quality scores and information-state observations), and selects actions by minimizing an expected free energy (EFE) objective. The authors report that the learned observation matrices develop 'structure,' that the EFE landscape differentiates policies over time, and that action selection shifts from search-dominated exploration to prompt-dominated exploitation. The paper includes a derivation of EFE, a description of the generative model and learning rules, and qualitative results based on visual inspection of heatmaps and timelines. A GitHub repository is cited as containing the full implementation.

Significance. If the central claims were established, the paper would offer a useful proof-of-concept that active inference can be applied to high-dimensional, language-driven agent control, with a principled objective balancing information gain and goal-seeking. The manuscript is honest about the fixed state-space limitation and releases code, which supports reproducibility. However, the significance is currently limited by two load-bearing problems: the EFE derivation in Section 3.3 relies on an incorrect probabilistic equality, and the empirical evaluation in Section 5 is entirely qualitative and self-referential, with no quantitative metrics, baselines, error bars, or external ground truth. The claimed 'accurate models of environment dynamics' and 'sophisticated exploration-exploitation behavior' are therefore not established by the evidence presented.

major comments (3)

[Section 3.3, Eqs. (10)-(14)] The derivation of the expected free energy is invalid. The asserted equality q(s|π)q(s|o,π) = q(o|π)q(o|s) in Eq. (10) is not Bayes' theorem; the correct identity is q(s|o,π)q(o|π) = q(o|s,π)q(s|π). Substituting the correct identity into the step from Eq. (12) to Eq. (13) yields terms involving q(o|π)q(s|o,π), not the expression the authors obtain. Consequently, Eqs. (13) and (14) do not follow, and the policy-selection objective Gπ used in Section 4.5 and Eq. (24) is not derived. Because the paper explicitly attributes the observed exploration-exploitation pattern to free-energy minimization, this invalid derivation undermines the central theoretical claim.
[Sections 4.4 and 5.1] The empirical claim that the agent learns 'accurate models of environment dynamics' is not supported by any external validation. The only source of observations is the GPT-4o-mini evaluator described in Section 4.4, which is part of the agent's own pipeline. The learned observation matrices in Figure 2 may simply mirror the evaluator's scoring tendencies rather than objective environment structure. There is no ground truth, no held-out prediction test, and no quantitative metric (e.g., predictive log-likelihood, calibration, or comparison to a random learner). Without such validation, the 'emergent structure' is not evidence of accurate environment modeling.
[Sections 4.5, 4.6, and 5.2-5.3] The reported exploration-exploitation transition is not demonstrated to emerge from free-energy minimization, because the preference matrix C (Section 4.2.3) strongly encodes a preference for 'detailed info' (value 64.0) over 'no info' (-32.0) and the action precision is high (α=16.0). These hand-set preferences, together with the invalid Gπ derivation, could trivially produce the observed shift from search to prompt actions. The paper provides no ablation, no sensitivity analysis, and no comparison against a baseline that uses random action selection or a simpler heuristic (e.g., always search until information state is high, then exploit). The claim that the behavior is 'sophisticated' and 'emerges from the framework' requires such controls.

minor comments (4)

[Algorithm 1, line 19] The notation `a_{t+1} <- a_t + η·(o_t ⊗ q(s_t)) ⊙ (A > 0)` appears to update the observation matrix A, not the action a; Eq. (22) uses `a_{t+1}^m` for Dirichlet concentration parameters. Please align the pseudocode with the mathematical notation.
[Figures 1-5] All figures lack axis labels, colorbar scales, and units. For example, Figure 3 shows EFE values but no quantitative scale, and Figure 5 uses 'prompt IDs' and 'search IDs' without defining the mapping to the 33 prompt and 11 search states. The qualitative claims would be more interpretable with labeled axes and a legend.
[References] Several references are malformed or incomplete: the 'graph websearch agent' GitHub entry contains a space in the repository name, the Chou/Champion et al. reference is incomplete, and two different works by Sun et al. (2023) share the same arXiv identifier. Please verify and correct the bibliography.
[Data Availability] The GitHub URL ends with 'Adaptat', which appears truncated; please provide the full, working URL and ideally a DOI or versioned archive so that the code can be accessed reproducibly.

Circularity Check

1 steps flagged · score 6.0 of 10

Partial circularity: the claimed 'accurate model of the environment' is the fitted observation matrix itself; action-selection claims are not circular but rely on an invalid EFE derivation that is a correctness risk.

fitted input called prediction [Sections 4.3.1 (Eq. 22), 4.4, and 5.1]
"Through active exploration and learning, the agent successfully developed an accurate model of the environment, particularly the relationships between states and observations. ... The agent receives structured observations about search quality and prompt effectiveness through evaluations performed by GPT-4o-mini ... This yields the learning rule: a_{t+1}^m = a_t^m + η · (o_m ⊗ q(s)) ⊙ (A_m > 0)"

The A matrices are updated by directly accumulating the evaluator's observations (Eq. 22), and Section 4.4 defines the 'environment' as the paper's own GPT-4o-mini evaluation pipeline. The final matrices shown in Figure 2 are therefore the empirical accumulation of those same observations. Presenting this structure as evidence of an 'accurate model of the environment' is a fitted value being reported as a validated result: the evidence (emergent structure) is the fit itself, with no held-out observations or external ground truth. The claim of accuracy reduces by construction to the learning rule that produced the matrices.

full rationale

The only step that reduces to its own inputs is the validation of the learned observation model. Eq. 22 updates the A matrices by directly accumulating observations, and Section 4.4 defines the environment as the GPT-4o-mini evaluation pipeline; Section 5.1 then treats the resulting structure as evidence of an 'accurate model of the environment.' That evidence is the fit itself, so this is a fitted input called a prediction. The exploration-exploitation claims are not circular in the same way: action selection minimizes the standard expected-free-energy objective (Eqs. 8-9) with hand-specified preferences (C matrix) and information-gain terms, and the temporal transition from search to prompt actions is a simulated emergent property, albeit strongly shaped by those inputs. There are no load-bearing self-citations (no prior work by the present author is cited), and no uniqueness theorem is imported. The derivation in Section 3.3 contains a serious mathematical error: Eq. 10 is asserted to hold for POMDPs but is not Bayes' theorem, so Eqs. 13-14 do not follow. This is a correctness risk, not a circularity, and the experiments use the earlier conceptual form (Eq. 8), so it does not raise the circularity score. Overall, one central validation claim is circular by construction, while the rest of the paper's derivation chain is independent of its inputs.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim depends on many hand-set parameters and assumptions. Most notably, the C prior and control parameters directly produce the reported exploration-exploitation pattern, and the 'environment' is an LLM evaluator whose reliability is assumed, not shown.

free parameters (8)

Learning rate eta = 50.0
Chosen ad hoc; controls the speed of Dirichlet parameter updates in Eq. 22.
Policy precision gamma = 8.0
Chosen ad hoc to balance exploration and exploitation in Eq. 24.
Action precision alpha = 16.0
Chosen ad hoc to control action selection determinism.
Preference matrix C (quality metrics) = -16 for low quality, quadratic scaling 2.0
Hand-specified to drive the agent toward high-quality observations; directly shapes the exploration-exploitation behavior.
Preference matrix C (information states) = -32.0, 8.0, 64.0
Hand-specified to strongly prefer detailed information; biases the agent toward information-seeking early on.
Dirichlet base concentration = 1.0
Chosen to maximize learning flexibility; standard but arbitrary.
Transition bias = 0.1
Small bias for specific transitions (persistence, decay, forward progression); set by hand.
Policy horizon and inference horizon = 2 and 1
Chosen without sensitivity analysis.

assumptions (6)

standard math Variational free energy bounds surprise via Jensen's inequality
Used in Section 3.2 to derive the VFE bound; standard.
domain assumption The environment is a POMDP with the specified state factors and observation modalities
Section 4.1 assumes prompt, search, and information factors fully explain observations; no test of this factorization.
domain assumption Expected free energy is a valid objective for action selection
Standard active inference assumption, cited to Friston et al.; not independently proven here.
domain assumption GPT-4o-mini evaluations provide reliable, calibrated quality scores
Section 4.4 relies on LLM-generated JSON scores as ground truth observations; no calibration or validation is reported.
ad hoc to paper The equality q(s|pi)q(s|o,pi) = q(o|pi)q(o|s) (Eq. 10) holds
Stated as a re-arrangement of Bayes' theorem, but it is not generally true and does not follow from the displayed equations; the derivation of Eq. 14 depends on it.
ad hoc to paper Thermodynamic costs (Jarzynski equality) apply to belief updating in this system
Invoked in Section 1 and 5.3 but never defined, measured, or used in the experiments; a rhetorical claim rather than an operational one.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Active Inference for Self-Organizing Multi-LLM Systems: A Bayesian Thermodynamic Approach to Adaptation." pith.science (2026). https://pith.science/paper/XFJDEBFS

@misc{pith2026241210425,
  author       = {Pith},
  title        = {Pith review of: Active Inference for Self-Organizing Multi-LLM Systems: A Bayesian Thermodynamic Approach to Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XFJDEBFS}},
  note         = {Machine review of arXiv:2412.10425}
}

read the original abstract

This paper introduces a novel approach to creating adaptive language agents by integrating active inference with large language models (LLMs). While LLMs demonstrate remarkable capabilities, their reliance on static prompts limits adaptation to new information and changing environments. We address this by implementing an active inference framework that acts as a cognitive layer above an LLM-based agent, dynamically adjusting prompts and search strategies through principled information-seeking behavior. Our framework models the environment using three state factors (prompt, search, and information states) with seven observation modalities capturing quality metrics. By framing the agent's learning through the free energy principle, we enable systematic exploration of prompt combinations and search strategies. Experimental results demonstrate the effectiveness of this approach, with the agent developing accurate models of environment dynamics evidenced by emergent structure in observation matrices. Action selection patterns reveal sophisticated exploration-exploitation behavior, transitioning from initial information-gathering to targeted prompt testing. The integration of thermodynamic principles with language model capabilities provides a principled framework for creating robust, adaptable agents, extending active inference beyond traditional low-dimensional control problems to high-dimensional, language-driven environments.

Figures

Figures reproduced from arXiv: 2412.10425 by the authors.

**Figure 2.** Final learned observation mappings after environment interaction. The first three matrices on the [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Progression of Expected Free Energy (EFE) values for different policies across four time points. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Heatmap showing the frequency of action selection across prompt and search dimensions. Lighter [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Time series of action selection throughout the experiment. Blue dots represent prompt actions [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Orchestrator: Active Inference for Multi-Agent Systems in Long-Horizon Tasks
cs.MA 2025-09 conditional novelty 4.0 of 10

Orchestrator, an active-inference-inspired feedback system for LLM multi-agent teams, substantially raises maze-solving success rates on medium-difficulty mazes but not consistently on hard mazes.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages · cited by 1 Pith paper

[1]

Adeojo, J. (2024). graph websearch agent: Websearch agent built on the LangGraph framework. GitHub repository. Retrieved from https://github.com/john-adeojo/graph websearch agent

work page 2024
[2]

Bai, Y., Saunders, W., Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 17117–17130

work page 2022
[3]

R., Deng, S., Raffel, C., et al

Bowman, S. R., Deng, S., Raffel, C., et al. (2022). Measuring progress on scalable oversight for large language models. arXiv preprint arXiv:2202.07765

arXiv 2022
[4]

R., & Friston, K

Brown, H. R., & Friston, K. J. (2018). The Physics of Free Will. Neuroscience and Biobehavioral Reviews, 90, 54–64. https://www.sciencedirect.com/science/article/pii/S0149763418302525?ref=pdf download&fr=RR- 2&rr=88d83f2d8dfbf279#bib0145

work page 2018
[5]

L., Kim, C

Buckley, C. L., Kim, C. S., McGregor, S., & Seth, A. K. (2017). The free energy principle for action and perception: A mathematical review. Biological Cybernetics, 112(6), 1–18. https://link.springer.com/article/10.1007/s00422-019-00805-w

doi:10.1007/s00422-019-00805-w 2017
[6]

Champion, T., Bowman, H., Markovi´ c, D., & Grze´ s, M. (2023). Reframing the Expected Free Energy: Four Formulations and a Unification. University of Kent, School of Computing, Canterbury, United Kingdom; University of Birmingham, School of Psychology and School of Computer Science,

work page 2023
[7]

Dandoy, L., & Di Francesco, M. (2023). Active Inference with State-Only Control. https://arxiv.org/pdf/2311.10300

arXiv 2023
[8]

F., Hazan, H., Levin, M., & Marcian` o, A

Fields, C., Fabrocini, F., Friston, K., Glazebrook, J. F., Hazan, H., Levin, M., & Marcian` o, A. (2023). Control flow in active inference systems. Allen Discovery Center at Tufts University. https://arxiv.org/abs/2303.03347

arXiv 2023

Show all 26 references

[9]

Friston, K., Parr, T., & de Vries, B. (2017). The graphical brain: Belief propagation and active inference. Network Neuroscience, 1(4), 381–414. https://doi.org/10.1162/NETN a 00018

2017 doi
[10]

Gou, W., Sun, X., Li, Q., et al. (2023a). Leveraging external tools for critique-driven self-improvement in language models. arXiv preprint arXiv:2306.05123

2023 arXiv
[11]

Guo, J., Liu, Y., Chen, W., et al. (2024). Re-ReST: Reflection-reinforced self-training for language agents. arXiv preprint arXiv:2403.07125

2024 arXiv
[12]

Liu, S., Li, Y., Zhang, K., et al. (2023). Odyssey: Empowering Minecraft agents with open-world skills. arXiv preprint arXiv:2310.01234

2023 arXiv
[14]

Millidge, B., Tschantz, A., & Buckley, C. L. (2020). Predictive Coding: A Theoretical and Experimental Review. NeurIPS. https://papers.nips.cc/paper files/paper/2020/file/865dfbde8a344b44095495f3591f7407-Paper.pdf

2020
[15]

Nascimento, N., Alencar, P., Cowan, D., et al. (2024). Generative AI for self-adaptive systems: State of the art and research roadmap. ACM Transactions on Autonomous and Adaptive Systems, 19(3), 1–60

2024
[16]

Parr, T., Pezzulo, G., & Friston, K. J. (2021). Active inference: The free energy principle in mind, brain, and behavior. Journal of Mathematical Psychology, 100, 102364. https://www.sciencedirect.com/science/article/pii/S0022249621000973#b40

2021
[17]

Sajid, N., Friston, K., & Parr, T. (2021). Planning and Active Inference. https://arxiv.org/abs/2103.13860v3

2021
[18]

Schwartenbeck, P., & Friston, K. (2017). Active Inference, Curiosity and Insight. Neural Computation, 29(10), 2633–2683. https://direct.mit.edu/neco/article-abstract/29/10/2633/8300/Active-Inference- Curiosity-and-Insight?redirectedFrom=fulltext

2017
[19]

Schwartenbeck, P., FitzGerald, T., Mathys, C., Dolan, R., & Friston, K. (2023). Active inference, belief propagation, and the free energy principle. PLOS ONE, 17(11), e0277199. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0277199

2023 doi
[20]

Shipp, S. (2023). The role of the free energy principle in cognitive systems. Cognitive Science, 17(2), 212–248. https://journals.sagepub.com/doi/pdf/10.1177/26339137231222481 13

2023 doi
[21]

Smith, J., & Johnson, M. (2023). Advances in Active Inference. Trends in Cognitive Sciences. https://www.sciencedirect.com/science/article/pii/S1364661323002607

2023
[22]

J., & Whyte, C

Smith, R., Friston, K. J., & Whyte, C. J. (2023). A step-by-step tutorial on active inference and its application to empirical data. Journal of Mathematical Psychology, 107, 102632

2023
[23]

Sun, Z., Wang, L., Li, Y., et al. (2023). Toward self-improvement of LLMs via imagination, searching, and criticizing. arXiv preprint arXiv:2310.00533

2023 arXiv
[24]

Wang, G., Xie, Y., Jiang, Y., et al. (2023). Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2306.07291

2023
[25]

Wang, P., Chen, J., Zhao, H., et al. (2023b). WEBRL: Training LLM web agents via self-evolving online curriculum reinforcement learning. arXiv preprint arXiv:2305.09876

2023 arXiv
[26]

Wortsman, M., Ehsani, K., Rastegari, M., et al. (2019). Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 6757–6765

2019
[27]

Zhang, Y., & Zheng, Z. (2023). Reinforcement Learning with Active Inference. https://arxiv.org/pdf/2306.09205 14

2023 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Adeojo, J. (2024). graph websearch agent: Websearch agent built on the LangGraph framework. GitHub repository. Retrieved from https://github.com/john-adeojo/graph websearch agent

work page 2024

[2] [2]

Bai, Y., Saunders, W., Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 17117–17130

work page 2022

[3] [3]

R., Deng, S., Raffel, C., et al

Bowman, S. R., Deng, S., Raffel, C., et al. (2022). Measuring progress on scalable oversight for large language models. arXiv preprint arXiv:2202.07765

arXiv 2022

[4] [4]

R., & Friston, K

Brown, H. R., & Friston, K. J. (2018). The Physics of Free Will. Neuroscience and Biobehavioral Reviews, 90, 54–64. https://www.sciencedirect.com/science/article/pii/S0149763418302525?ref=pdf download&fr=RR- 2&rr=88d83f2d8dfbf279#bib0145

work page 2018

[5] [5]

L., Kim, C

Buckley, C. L., Kim, C. S., McGregor, S., & Seth, A. K. (2017). The free energy principle for action and perception: A mathematical review. Biological Cybernetics, 112(6), 1–18. https://link.springer.com/article/10.1007/s00422-019-00805-w

doi:10.1007/s00422-019-00805-w 2017

[6] [6]

Champion, T., Bowman, H., Markovi´ c, D., & Grze´ s, M. (2023). Reframing the Expected Free Energy: Four Formulations and a Unification. University of Kent, School of Computing, Canterbury, United Kingdom; University of Birmingham, School of Psychology and School of Computer Science,

work page 2023

[7] [7]

Dandoy, L., & Di Francesco, M. (2023). Active Inference with State-Only Control. https://arxiv.org/pdf/2311.10300

arXiv 2023

[8] [8]

F., Hazan, H., Levin, M., & Marcian` o, A

Fields, C., Fabrocini, F., Friston, K., Glazebrook, J. F., Hazan, H., Levin, M., & Marcian` o, A. (2023). Control flow in active inference systems. Allen Discovery Center at Tufts University. https://arxiv.org/abs/2303.03347

arXiv 2023