arxiv: 2604.07745 · v1 · submitted 2026-04-09 · 💻 cs.AI · q-bio.NC

Recognition: no theorem link

The Cartesian Cut in Agentic AI

Tim Sainburg , Caleb Weinreb

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:52 UTC · model grok-4.3

classification 💻 cs.AI q-bio.NC

keywords LLM agentsCartesian agencycontrol architecturefeedback controllersautonomyrobustnessoversightagentic AI

0 comments

The pith

LLM agents implement Cartesian agency by coupling a learned predictive core to an external engineered runtime via a symbolic interface.

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

The paper claims that where control resides is a central design lever in agentic AI systems. Biological brains embed prediction within layered feedback controllers that are calibrated directly by the outcomes of actions. By contrast, LLM agents turn text prediction into control by linking the model to a separate runtime that handles state and policies through symbolic connections. This Cartesian split supports bootstrapping from existing language models, modular development, and external governance, yet it can create interface sensitivities and control bottlenecks. The authors contrast bounded services, Cartesian agents, and integrated agents to illustrate resulting trade-offs in autonomy, robustness, and oversight.

Core claim

LLMs gain competence by predicting words in human text, which often reflects how people perform tasks. Consequently, coupling an LLM to an engineered runtime turns prediction into control: outputs trigger interventions that enact goal-oriented behavior. Brains embed prediction within layered feedback controllers calibrated by the consequences of action. By contrast, LLM agents implement Cartesian agency: a learned core coupled to an engineered runtime via a symbolic interface that externalizes control state and policies. The split enables bootstrapping, modularity, and governance, but can induce sensitivity and bottlenecks.

What carries the argument

Cartesian agency, the mechanism in which a learned predictive core is coupled to an engineered runtime through a symbolic interface that externalizes control state and policies.

If this is right

The symbolic interface allows competence from text prediction to be turned into goal-directed interventions without retraining the core model.
Externalizing control state and policies supports modularity and easier oversight or governance of the agent's behavior.
The separation can create bottlenecks at the interface and increase sensitivity to mismatches between the learned predictions and the runtime's needs.
Bounded services limit the scope of autonomous action to improve robustness, while integrated agents reduce the split to increase autonomy at the cost of harder oversight.

Where Pith is reading between the lines

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

Designers could test whether reducing the symbolic interface in favor of tighter prediction-control coupling improves performance on tasks with frequent environmental changes.
The framework suggests hybrid architectures that keep some external governance while embedding more feedback calibration inside the learned component.
Comparing failure modes across the three approaches on long-horizon tasks would clarify which trade-off dominates in practice.

Load-bearing premise

The location of control is the central design lever, and the described split between learned core and external runtime accurately captures the functional differences from biological feedback systems.

What would settle it

An experiment that measures robustness and autonomy metrics when the same task is solved by an LLM agent using the standard symbolic interface versus an otherwise identical system that integrates prediction directly into feedback control loops.

Figures

Figures reproduced from arXiv: 2604.07745 by Caleb Weinreb, Tim Sainburg.

read the original abstract

LLMs gain competence by predicting words in human text, which often reflects how people perform tasks. Consequently, coupling an LLM to an engineered runtime turns prediction into control: outputs trigger interventions that enact goal-oriented behavior. We argue that a central design lever is where control resides in these systems. Brains embed prediction within layered feedback controllers calibrated by the consequences of action. By contrast, LLM agents implement Cartesian agency: a learned core coupled to an engineered runtime via a symbolic interface that externalizes control state and policies. The split enables bootstrapping, modularity, and governance, but can induce sensitivity and bottlenecks. We outline bounded services, Cartesian agents, and integrated agents as contrasting approaches to control that trade off autonomy, robustness, and oversight.

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.

Desk Editor's Note private letter to a colleague

The paper frames LLM agents as Cartesian systems with a learned core separated from engineered control, which organizes some design trade-offs but stays conceptual without new evidence.

read the letter

The main point is that current LLM agents keep prediction inside the model but move control and state into an external runtime connected by a symbolic interface, unlike brains that embed prediction inside layered feedback loops calibrated by action outcomes. This split is presented as a central design lever that buys modularity, easier bootstrapping, and oversight at the cost of possible sensitivity and bottlenecks. The authors sketch three approaches—bounded services, Cartesian agents, and integrated agents—to show how different placements of control trade off autonomy, robustness, and governance. That taxonomy is the clearest part of the work and gives a practical way to talk about where to draw the line in agent architectures. The biological contrast is drawn at a high level but helps highlight what is distinctive about the LLM setup. The soft spots are straightforward: the whole argument is observational and analogical. There are no experiments, no quantitative results, no formal derivations, and no direct comparisons that would show how large the claimed trade-offs actually are in practice. The core separation of learning from control has appeared in earlier hybrid AI work, so the novelty is mostly in applying the Cartesian label and the specific contrasts here. The assumption that control location is the dominant lever is stated but not tested against other factors like model scale or interface design. This is for people building or reviewing agent systems who want a compact framework for discussing governance and robustness choices. A reader already familiar with agentic AI literature will recognize the intuitions but may still find the organization useful. It deserves peer review as a perspective piece because the framing is coherent and the topic matters, though any referee would likely ask for concrete examples or small tests to ground the trade-offs.

Referee Report

1 major / 2 minor

Summary. The paper claims that a central design lever in agentic AI systems is the location of control. Biological brains embed prediction within layered feedback controllers calibrated by the consequences of action, whereas LLM agents implement Cartesian agency: a learned predictive core coupled to an engineered runtime via a symbolic interface that externalizes control state and policies. This split enables bootstrapping, modularity, and governance but induces sensitivity and bottlenecks. The manuscript outlines three contrasting approaches—bounded services, Cartesian agents, and integrated agents—and analyzes their trade-offs in autonomy, robustness, and oversight.

Significance. If the framing holds, the paper supplies a coherent interpretive lens for analyzing architectural choices in agentic AI by centering control location and deriving logical consequences of the posited Cartesian split from current LLM usage patterns and biological descriptions. It earns credit for presenting the distinction clearly and mapping trade-offs directly to the three outlined approaches without hidden parameters or circular definitions.

major comments (1)

Abstract: the central claim that the symbolic interface externalizes control state and policies is load-bearing for all subsequent trade-off analysis, yet the manuscript provides no explicit criteria or examples for identifying such an interface in deployed LLM agents (e.g., tool-calling loops or memory buffers), leaving the distinction at a level of generality that limits falsifiability and concrete application.

minor comments (2)

The title invokes a 'Cartesian Cut' that is not formally defined or situated relative to prior uses of the term in philosophy or systems theory; a brief clarifying sentence would improve accessibility.
The biological comparison would benefit from one or two additional citations to specific control-theoretic or neuroscientific models of layered feedback to strengthen the contrast without altering the conceptual nature of the argument.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment of the manuscript's framing and significance, as well as the recommendation for minor revision. We address the single major comment below.

read point-by-point responses

Referee: Abstract: the central claim that the symbolic interface externalizes control state and policies is load-bearing for all subsequent trade-off analysis, yet the manuscript provides no explicit criteria or examples for identifying such an interface in deployed LLM agents (e.g., tool-calling loops or memory buffers), leaving the distinction at a level of generality that limits falsifiability and concrete application.

Authors: We agree that greater concreteness in the abstract would strengthen the load-bearing claim and improve applicability. In the revised manuscript we will expand the abstract to include explicit examples and identification criteria, such as tool-calling loops (where LLM token predictions are parsed into structured external function calls that update runtime state and policies) and memory buffers (where control policies and state are maintained outside the predictive core). These additions will ground the Cartesian split in observable architectural patterns without changing the core arguments or trade-off analysis. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper advances a conceptual perspective framing control location as a design lever and contrasting embedded biological feedback with LLM Cartesian agency via symbolic interfaces. No equations, fitted parameters, quantitative predictions, or deductive derivations are present that could reduce to inputs by construction. The argument rests on observational descriptions of existing systems and logical consequences of the posited split, without self-definitional loops, self-citation load-bearing premises, or renamed empirical patterns. The core claims remain self-contained as an interpretive framework.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on domain assumptions about LLM training and brain function plus a newly introduced conceptual distinction; no free parameters or quantitative fitting are present.

axioms (2)

domain assumption LLMs gain competence by predicting words in human text, which often reflects how people perform tasks.
Foundational premise stated in the abstract for how LLMs acquire task-related competence.
domain assumption Brains embed prediction within layered feedback controllers calibrated by the consequences of action.
Biological premise used to contrast with LLM agent architecture.

invented entities (1)

Cartesian agency no independent evidence
purpose: To label the split between learned predictive core and engineered runtime in LLM agents.
New term introduced to frame the design choice and its consequences.

pith-pipeline@v0.9.0 · 5411 in / 1413 out tokens · 36193 ms · 2026-05-10T17:52:08.909356+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

68 extracted references · 47 canonical work pages · 20 internal anchors

[1]

Parallel organization of functionally segregated circuits linking basal ganglia and cortex

Garrett E. Alexander, Mahlon R. DeLong, and Peter L. Strick. “Parallel organization of functionally segregated circuits linking basal ganglia and cortex”. In:Annual Review of Neuroscience9.1 (1986), pp. 357–381.doi:10.1146/annurev.ne.09.030186.002041. url:https://doi.org/10.1146/annurev.ne.09.030186.002041

work page doi:10.1146/annurev.ne.09.030186.002041 1986
[2]

Concrete Problems in AI Safety

Dario Amodei et al. “Concrete Problems in AI Safety”. In:arXiv(2016).doi:10. 48550/arXiv.1606.06565. arXiv:1606.06565 [cs.AI].url:https://arxiv.org/ abs/1606.06565

work page internal anchor Pith review arXiv 2016
[3]

Thinking inside the box: controlling and using an Oracle AI

Stuart Armstrong, Anders Sandberg, and Nick Bostrom. “Thinking inside the box: controlling and using an Oracle AI”. In:Minds and Machines22.4 (2012), pp. 299– 324.doi:10.1007/s11023-012-9282-2.url:https://doi.org/10.1007/s11023- 012-9282-2

work page doi:10.1007/s11023-012-9282-2.url:https://doi.org/10.1007/s11023- 2012
[4]

Self-supervised learning from images with a joint-embedding predictive architecture.arXiv preprint arXiv:2301.08243,

Mahmoud Assran et al. “Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture”. In:arXiv preprint arXiv:2301.08243(2023).doi:10.48550/ arXiv.2301.08243

work page arXiv 2023
[5]

V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

Mido Assran et al. “V-JEPA 2: Self-Supervised Video Models Enable Understand- ing, Prediction and Planning”. In:arXiv preprint arXiv:2506.09985(2025).doi:10. 48550/arXiv.2506.09985

work page internal anchor Pith review arXiv 2025
[6]

An integrative theory of locus coeruleus- norepinephrine function: adaptive gain and optimal performance

Gary Aston-Jones and Jonathan D. Cohen. “An integrative theory of locus coeruleus- norepinephrine function: adaptive gain and optimal performance”. In:Annual Review of Neuroscience28 (2005), pp. 403–450.doi:10.1146/annurev.neuro.28.061604. 135709.url:https://doi.org/10.1146/annurev.neuro.28.061604.135709

work page doi:10.1146/annurev.neuro.28.061604 2005
[7]

Working paper / preprint (under review, per publisher page at time of access)

Fazl Barez et al.Chain-of-Thought Is Not Explainability. Working paper / preprint (under review, per publisher page at time of access). 2025.url:https://www.aigi. ox.ac.uk/publications/chain-of-thought-is-not-explainability/(visited on 01/22/2026)

2025
[8]

Canonical microcircuits for predictive coding

Andre M. Bastos et al. “Canonical microcircuits for predictive coding”. In:Neuron 76.4 (2012), pp. 695–711.doi:10 . 1016 / j . neuron . 2012 . 10 . 038.url:https : //doi.org/10.1016/j.neuron.2012.10.038

work page doi:10.1016/j.neuron.2012.10.038 2012
[9]

Climbing towards NLU: On meaning, form, and understanding in the age of data

Emily M Bender and Alexander Koller. “Climbing towards NLU: On meaning, form, and understanding in the age of data”. In:Proceedings of the 58th annual meeting of the association for computational linguistics. 2020, pp. 5185–5198. 18

2020
[10]

arXiv preprint arXiv:2502.15657 , year=

Yoshua Bengio et al. “Superintelligent agents pose catastrophic risks: Can scientist AI offer a safer path?” In:arXiv(2025).doi:10.48550/arXiv.2502.15657. arXiv: 2502.15657 [cs.AI].url:https://arxiv.org/abs/2502.15657

work page doi:10.48550/arxiv.2502.15657 2025
[11]

On the Opportunities and Risks of Foundation Models

Rishi Bommasani et al. “On the Opportunities and Risks of Foundation Models”. In: arXiv(2021).doi:10.48550/arXiv.2108.07258. arXiv:2108.07258 [cs.LG].url: https://arxiv.org/abs/2108.07258

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2108.07258 2021
[12]

Oxford University Press, 2014.isbn: 9780199678112.url:https://www.oxfordmartin.ox.ac.uk/publications/ superintelligence-paths-dangers-strategies/(visited on 01/22/2026)

Nick Bostrom.Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.isbn: 9780199678112.url:https://www.oxfordmartin.ox.ac.uk/publications/ superintelligence-paths-dangers-strategies/(visited on 01/22/2026)

2014
[13]

Language Models are Few-Shot Learners

Tom B. Brown et al. “Language Models are Few-Shot Learners”. In:Advances in Neural Information Processing Systems33 (2020), pp. 1877–1901.doi:10 . 48550 / arXiv.2005.14165. arXiv:2005.14165 [cs.CL].url:https://arxiv.org/abs/ 2005.14165

work page internal anchor Pith review Pith/arXiv arXiv 2020
[14]

Sparks of Artificial General Intelligence: Early experiments with GPT-4

S´ ebastien Bubeck et al. “Sparks of artificial general intelligence: Early experiments with gpt-4”. In:arXiv preprint arXiv:2303.12712(2023)

work page internal anchor Pith review arXiv 2023
[15]

2022.url:https : / / github

Harrison Chase.LangChain. 2022.url:https : / / github . com / langchain - ai / langchain

2022
[16]

Does AI already have human-level intelligence? The evi- dence is clear

Eddy Keming Chen et al. “Does AI already have human-level intelligence? The evi- dence is clear”. In:Nature650 (2026), pp. 36–40.doi:10.1038/d41586-026-00285-6. url:https://www.nature.com/articles/d41586-026-00285-6

work page doi:10.1038/d41586-026-00285-6 2026
[17]

Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks

Wenhu Chen et al. “Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks”. In:arXiv(2022). Also published in TMLR (2023).doi:10.48550/arXiv.2211.12588. arXiv:2211.12588 [cs.CL].url: https://arxiv.org/abs/2211.12588

work page internal anchor Pith review doi:10.48550/arxiv.2211.12588 2022
[18]

Deep reinforcement learning from human preferences

Paul F Christiano et al. “Deep reinforcement learning from human preferences”. In: Advances in neural information processing systems30 (2017)

2017
[19]

Cortical mechanisms of action selection: the affordance competition hy- pothesis

Paul Cisek. “Cortical mechanisms of action selection: the affordance competition hy- pothesis”. In:Philosophical Transactions of the Royal Society B: Biological Sciences 362.1485 (2007), pp. 1585–1599.doi:10.1098/rstb.2007.2054.url:https://doi. org/10.1098/rstb.2007.2054

work page doi:10.1098/rstb.2007.2054.url:https://doi 2007
[20]

Evolution of behavioural control from chordates to primates

Paul Cisek. “Evolution of behavioural control from chordates to primates”. In:Philo- sophical Transactions of the Royal Society B: Biological Sciences377.1844 (2022), p. 20200522.doi:10.1098/rstb.2020.0522.url:https://doi.org/10.1098/ rstb.2020.0522. 19

work page doi:10.1098/rstb.2020.0522.url:https://doi.org/10.1098/ 2022
[21]

The extended mind

Andy Clark and David Chalmers. “The extended mind”. In:analysis58.1 (1998), pp. 7–19

1998
[22]

How Do You Feel? Interoception: the sense of the physiological condition of the body

A. D. (Bud) Craig. “How Do You Feel? Interoception: the sense of the physiological condition of the body”. In:Nature Reviews Neuroscience3.8 (2002), pp. 655–666.doi: 10.1038/nrn894.url:https://doi.org/10.1038/nrn894

work page doi:10.1038/nrn894.url:https://doi.org/10.1038/nrn894 2002
[23]

Eric Drexler.Reframing Superintelligence: Comprehensive AI Services as General Intelligence

K. Eric Drexler.Reframing Superintelligence: Comprehensive AI Services as General Intelligence. Tech. rep. Future of Humanity Institute, University of Oxford, 2019.url: https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence. pdf(visited on 01/22/2026)

2019
[24]

ReTool: Reinforcement Learning for Strategic Tool Use in LLMs

Jiazhan Feng et al. “ReTool: Reinforcement Learning for Strategic Tool Use in LLMs”. In:International Conference on Learning Representations (ICLR). Poster. 2026.url: https://openreview.net/forum?id=tRk1nofSmz

2026
[25]

A theory of cortical responses.Philosophical Transactions of the Royal Society B: Biological Sciences, 360(1456):815–836, 2005

Karl Friston. “A theory of cortical responses”. In:Philosophical Transactions of the Royal Society B: Biological Sciences360.1456 (2005), pp. 815–836.doi:10 . 1098 / rstb.2005.1622.url:https://doi.org/10.1098/rstb.2005.1622

work page doi:10.1098/rstb.2005.1622 2005
[26]

PAL: Program-aided Language Models

Luyu Gao et al. “PAL: Program-aided Language Models”. In:arXiv(2022).doi:10. 48550/arXiv.2211.10435. arXiv:2211.10435 [cs.CL].url:https://arxiv.org/ abs/2211.10435

work page Pith review arXiv 2022
[27]

Gemini Robotics: Bringing AI into the Physical World

Gemini Robotics Team. “Gemini Robotics: Bringing AI into the Physical World”. In: arXiv preprint arXiv:2503.20020(2025).doi:10.48550/arXiv.2503.20020

work page internal anchor Pith review doi:10.48550/arxiv.2503.20020 2025
[28]

Alignment faking in large language models

Ryan Greenblatt et al. “Alignment faking in large language models”. In:arXiv preprint arXiv:2412.14093(2024)

work page internal anchor Pith review arXiv 2024
[29]

Central Pattern Generators for Locomotion, with Spe- cial Reference to Vertebrates

Sten Grillner and Peter Wall´ en. “Central Pattern Generators for Locomotion, with Spe- cial Reference to Vertebrates”. In:Annual Review of Neuroscience8 (1985), pp. 233– 261.doi:10.1146/annurev.ne.08.030185.001313.url:https://doi.org/10. 1146/annurev.ne.08.030185.001313

work page doi:10.1146/annurev.ne.08.030185.001313.url:https://doi.org/10 1985
[30]

Mastering Atari with Discrete World Models

Danijar Hafner et al. “Mastering Atari with discrete world models”. In:arXiv(2020). doi:10 . 48550 / arXiv . 2010 . 02193. arXiv:2010 . 02193 [cs.LG].url:https : / / arxiv.org/abs/2010.02193

work page internal anchor Pith review arXiv 2020
[31]

Language Models (Mostly) Know What They Know

Saurav Kadavath, Tom Conerly, Amanda Askell, et al. “Language Models (Mostly) Know What They Know”. In:arXiv preprint arXiv:2207.05221(2022).doi:10.48550/ arXiv.2207.05221. 20

work page internal anchor Pith review Pith/arXiv arXiv 2022
[32]

Kandel et al.Principles of Neural Science

Eric R. Kandel et al.Principles of Neural Science. 6th ed. New York, NY: McGraw Hill, 2021.url:https : / / accessbiomedicalscience . mhmedical . com / content . aspx?aid=1180370208(visited on 01/22/2026)

2021
[33]

Scaling Laws for Neural Language Models

Jared Kaplan et al. “Scaling Laws for Neural Language Models”. In:arXiv(2020). doi:10 . 48550 / arXiv . 2001 . 08361. arXiv:2001 . 08361 [cs.LG].url:https : / / arxiv.org/abs/2001.08361

work page internal anchor Pith review Pith/arXiv arXiv 2020
[34]

The ecological roles of bacterial chemotaxis

Johannes M Keegstra, Francesco Carrara, and Roman Stocker. “The ecological roles of bacterial chemotaxis”. In:Nature Reviews Microbiology20.8 (2022), pp. 491–504

2022
[35]

OpenVLA: An Open-Source Vision-Language-Action Model

Moo Jin Kim et al. “OpenVLA: An Open-Source Vision-Language-Action Model”. In: arXiv preprint arXiv:2406.09246(2024).doi:10.48550/arXiv.2406.09246

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.09246 2024
[36]

Version 0.9.2 (2022- 06-27)

Yann LeCun.A Path Towards Autonomous Machine Intelligence. Version 0.9.2 (2022- 06-27). 2022.url:https : / / openreview . net / pdf ? id = BZ5a1r - kVsf(visited on 01/22/2026)

2022
[37]

Universal Intelligence: A Definition of Machine In- telligence

Shane Legg and Marcus Hutter. “Universal Intelligence: A Definition of Machine In- telligence”. In:Minds and Machines17.4 (2007), pp. 391–444.doi:10.1007/s11023- 007-9079-x.url:https://doi.org/10.1007/s11023-007-9079-x

work page doi:10.1007/s11023- 2007
[38]

Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems

Sergey Levine et al. “Offline Reinforcement Learning: Tutorial, Review, and Perspec- tives on Open Problems”. In:arXiv(2020).doi:10.48550/arXiv.2005.01643. arXiv: 2005.01643 [cs.LG].url:https://arxiv.org/abs/2005.01643

work page internal anchor Pith review doi:10.48550/arxiv.2005.01643 2020
[39]

Let’s verify step by step

Hunter Lightman et al. “Let’s verify step by step”. In:The Twelfth International Conference on Learning Representations. 2023

2023
[40]

Lost in the middle: How language models use long contexts

Nelson F Liu et al. “Lost in the middle: How language models use long contexts”. In: Transactions of the association for computational linguistics12 (2024), pp. 157–173

2024
[41]

Context-dependent computation by recurrent dynamics in pre- frontal cortex

Valerio Mante et al. “Context-dependent computation by recurrent dynamics in pre- frontal cortex”. In:Nature503 (2013), pp. 78–84.doi:10.1038/nature12742.url: https://doi.org/10.1038/nature12742

work page doi:10.1038/nature12742.url: 2013
[42]

A hierarchy of intrinsic timescales across primate cortex

John D. Murray et al. “A hierarchy of intrinsic timescales across primate cortex”. In:Nature Neuroscience17.12 (2014), pp. 1661–1663.doi:10.1038/nn.3862.url: https://doi.org/10.1038/nn.3862

work page doi:10.1038/nn.3862.url: 2014
[43]

WebGPT: Browser-assisted question-answering with human feedback

Reiichiro Nakano et al. “WebGPT: Browser-assisted question-answering with human feedback”. In:arXiv(2021).doi:10.48550/arXiv.2112.09332. arXiv:2112.09332 [cs.CL].url:https://arxiv.org/abs/2112.09332. 21

work page internal anchor Pith review doi:10.48550/arxiv.2112.09332 2021
[44]

2024.url:https : / / www

Richard Ngo.The Agency Overhang. 2024.url:https : / / www . lesswrong . com / posts/tqs4eEJapFYSkLGfR/the-agency-overhang(visited on 03/27/2026)

2024
[45]

Training language models to follow instructions with human feedback

Long Ouyang et al. “Training language models to follow instructions with human feedback”. In:Advances in neural information processing systems35 (2022), pp. 27730– 27744

2022
[46]

Humans and automation: Use, misuse, disuse, abuse

Raja Parasuraman and Victor Riley. “Humans and automation: Use, misuse, disuse, abuse”. In:Human factors39.2 (1997), pp. 230–253

1997
[47]

Generating meaning: active inference and the scope and limits of passive AI

Giovanni Pezzulo et al. “Generating meaning: active inference and the scope and limits of passive AI”. In:Trends in Cognitive Sciences28.2 (2024), pp. 97–112.doi:10.1016/ j.tics.2023.10.002.url:https://doi.org/10.1016/j.tics.2023.10.002

work page doi:10.1016/j.tics.2023.10.002 2024
[48]

Powers.Behavior: The Control of Perception

William T. Powers.Behavior: The Control of Perception. Chicago, IL: Aldine, 1973, pp. xi, 296.url:http://www.livingcontrolsystems.com/books/bcp/bcp.html (visited on 01/22/2026)

1973
[49]

ToolRL: Reward is All Tool Learning Needs

Cheng Qian et al. “ToolRL: Reward is All Tool Learning Needs”. In:arXiv preprint arXiv:2504.13958(2025). arXiv:2504.13958 [cs.AI]

work page internal anchor Pith review arXiv 2025
[50]

OpenAI, 2019.url:https : / / cdn

Alec Radford et al.Language Models are Unsupervised Multitask Learners. OpenAI, 2019.url:https : / / cdn . openai . com / better - language - models / language _ models_are_unsupervised_multitask_learners.pdf(visited on 01/22/2026)

2019
[51]

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

Rajesh P. N. Rao and Dana H. Ballard. “Predictive coding in the visual cortex: a func- tional interpretation of some extra-classical receptive-field effects”. In:Nature Neuro- science2.1 (1999), pp. 79–87.doi:10.1038/4580.url:https://doi.org/10.1038/ 4580

work page doi:10.1038/4580.url:https://doi.org/10.1038/ 1999
[52]

The basal ganglia: a vertebrate solution to the selection problem?

Peter Redgrave, Tony J. Prescott, and Kevin Gurney. “The basal ganglia: a vertebrate solution to the selection problem?” In:Neuroscience89.4 (1999), pp. 1009–1023.doi: 10 . 1016 / S0306 - 4522(98 ) 00319 - 4.url:https : / / doi . org / 10 . 1016 / S0306 - 4522(98)00319-4

1999
[53]

Prototyping next-generation O-RAN research testbeds with SDRs,

Scott Reed et al. “A Generalist Agent”. In:arXiv(2022).doi:10.48550/arXiv.2205. 06175. arXiv:2205.06175 [cs.LG].url:https://arxiv.org/abs/2205.06175

work page doi:10.48550/arxiv.2205 2022
[54]

A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning

St´ ephane Ross, Geoffrey J. Gordon, and J. Andrew Bagnell. “A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning”. In:Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AIS- TATS). 2011, pp. 627–635.doi:10 . 48550 / arXiv . 1011 . 0686. arXiv:1011 . 0686 [cs.LG].url:https:...

work page Pith review arXiv 2011
[55]

Toolformer: Language Models Can Teach Themselves to Use Tools

Timo Schick et al. “Toolformer: Language Models Can Teach Themselves to Use Tools”. In:arXiv(2023).doi:10 . 48550 / arXiv . 2302 . 04761. arXiv:2302 . 04761 [cs.CL].url:https://arxiv.org/abs/2302.04761

work page internal anchor Pith review arXiv 2023
[56]

doi:10.1038/s41586-020- 3001-6

Julian Schrittwieser et al. “Mastering Atari, Go, Chess and Shogi by planning with a learned model”. In:Nature588.7839 (2020), pp. 604–609.doi:10.1038/s41586-020- 03051-4.url:https://doi.org/10.1038/s41586-020-03051-4

work page doi:10.1038/s41586-020- 2020
[57]

Stains,et al., Anatomy of STEM Teaching in American Universities: A Snapshot from a Large-Scale Observation Study.Science359(6383), 1468–1470 (2018), doi:10.1126/science

Wolfram Schultz, Peter Dayan, and P. Read Montague. “A neural substrate of predic- tion and reward”. In:Science275.5306 (1997), pp. 1593–1599.doi:10.1126/science. 275.5306.1593.url:https://doi.org/10.1126/science.275.5306.1593

work page doi:10.1126/science 1997
[58]

Quantifying language models’ sensitivity to spurious features in prompt design.arXiv preprint arXiv:2310.11324, 2023

Melanie Sclar et al. “Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting”. In: International Conference on Learning Representations (ICLR). 2024.doi:10.48550/ arXiv.2310.11324. arXiv:2310.11324 [cs.CL].url:https://arxiv.org/abs/ 2310.11324

work page arXiv 2024
[59]

Defining and Characterizing Reward Gaming

Joar Skalse et al. “Defining and Characterizing Reward Gaming”. In:Advances in Neural Information Processing Systems. 2022.url:https://arxiv.org/abs/2209. 13085

2022
[60]

Sutton.The Bitter Lesson

Richard S. Sutton.The Bitter Lesson. 2019.url:http://www.incompleteideas. net/IncIdeas/BitterLesson.html(visited on 01/22/2026)

2019
[61]

Optimal feedback control as a theory of motor coordination,

Emanuel Todorov and Michael I. Jordan. “Optimal feedback control as a theory of motor coordination”. In:Nature Neuroscience5.11 (2002), pp. 1226–1235.doi:10. 1038/nn963.url:https://doi.org/10.1038/nn963

work page doi:10.1038/nn963 2002
[62]

How to make artificial agents more like natural agents

Michael Tomasello. “How to make artificial agents more like natural agents”. In:Trends in Cognitive Sciences29.9 (2025), pp. 783–786.doi:10.1016/j.tics.2025.07.004. url:https://doi.org/10.1016/j.tics.2025.07.004

work page doi:10.1016/j.tics.2025.07.004 2025
[63]

Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

Miles Turpin et al. “Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting”. In:arXiv(2023). NeurIPS 2023.doi: 10.48550/arXiv.2305.04388. arXiv:2305.04388 [cs.CL].url:https://arxiv. org/abs/2305.04388

work page internal anchor Pith review doi:10.48550/arxiv.2305.04388 2023
[64]

Emergent Abilities of Large Language Models

Jason Wei et al. “Emergent abilities of large language models”. In:arXiv preprint arXiv:2206.07682(2022)

work page internal anchor Pith review arXiv 2022
[65]

Internal models in the cerebel- lum

Daniel M. Wolpert, R. Chris Miall, and Mitsuo Kawato. “Internal models in the cerebel- lum”. In:Trends in Cognitive Sciences2.9 (1998), pp. 338–347.doi:10.1016/S1364- 6613(98)01221-2.url:https://doi.org/10.1016/S1364-6613(98)01221-2. 23

work page doi:10.1016/s1364- 1998
[66]

ReAct: Synergizing Reasoning and Acting in Language Models

Shunyu Yao et al. “ReAct: Synergizing Reasoning and Acting in Language Models”. In: International Conference on Learning Representations (ICLR). 2023.doi:10.48550/ arXiv.2210.03629. arXiv:2210.03629 [cs.CL].url:https://arxiv.org/abs/ 2210.03629

work page internal anchor Pith review Pith/arXiv arXiv 2023
[67]

Uncertainty, neuromodulation, and attention

Angela J. Yu and Peter Dayan. “Uncertainty, neuromodulation, and attention”. In: Neuron46.4 (2005), pp. 681–692.doi:10 . 1016 / j . neuron . 2005 . 04 . 026.url: https://doi.org/10.1016/j.neuron.2005.04.026

work page doi:10.1016/j.neuron.2005.04.026 2005
[68]

RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

Brianna Zitkovich et al. “RT-2: Vision-Language-Action Models Transfer Web Knowl- edge to Robotic Control”. In:Conference on Robot Learning. PMLR. 2023, pp. 2165– 2183.doi:10.48550/arXiv.2307.15818. arXiv:2307.15818 [cs.RO].url:https: //arxiv.org/abs/2307.15818. 24

work page internal anchor Pith review doi:10.48550/arxiv.2307.15818 2023