Divergent Minds, Convergent Baselines: A Bounded-Rationality Account of LLM-Human Strategic Behaviour

Po Han Teo

arxiv: 2605.26437 · v1 · pith:6EMTXBNHnew · submitted 2026-05-26 · 💰 econ.GN · q-fin.EC

Divergent Minds, Convergent Baselines: A Bounded-Rationality Account of LLM-Human Strategic Behaviour

Po Han Teo This is my paper

Pith reviewed 2026-07-01 16:53 UTC · model grok-4.3

classification 💰 econ.GN q-fin.EC

keywords bounded rationalityLLM agentsstrategic behaviordelta correctioncognitive hierarchybehavioral economicscomputational boundsgame theory

0 comments

The pith

LLMs reach rational play in canonical games by retrieving solutions from training corpora instead of incurring the computational bound that shifts human choices.

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

The paper models human strategic choices as an unbounded rational baseline plus an additive correction term δ that records the cost of bounded computation. It argues that LLMs avoid producing the same δ in familiar games because they can retrieve and recombine solutions already present in training data. This account explains persistent gaps that fine-tuning and persona methods have not closed. A sympathetic reader would care because the substitution of LLM agents for human subjects in experiments rests on the assumption that both face comparable limits.

Core claim

For canonical games whose solutions are present in standard training corpora, LLMs retrieve and recombine corpus material, bypassing the bound that produces δ in humans. The framing extends to reasoning-distilled models through cognitive-hierarchy theory: their accessible level-k strategic reasoning is bounded by compute budget and context length rather than by the cognitive constraints that bound humans, and the δ they produce, if any, carries different structural signatures.

What carries the argument

The δ term, read as the mathematical signature of bounded computation: the gap between what an unboundedly-rational agent would compute and what a computationally bounded agent actually produces.

If this is right

Four operational tests discriminate human-shaped δ from LLM-shaped δ.
The magnitude of δ scales with peer-signal individuation, producing Cohen's d at least 0.5 between named-opponent and aggregate-opponent conditions.
Reasoning-distilled models produce δ with different structural signatures from human δ because their bounds arise from compute budget and context length.
The substitution of LLMs for humans in strategic experiments will continue to fail under current alignment methods.

Where Pith is reading between the lines

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

In games whose solutions are absent from training data, LLMs should produce δ signatures closer to those observed in humans.
Experiments could be redesigned around novel or paraphrased games to reduce corpus leakage when LLMs are used as subject proxies.
The moderator prediction supplies a quantitative threshold that could be checked directly in any decision environment that varies opponent identifiability.

Load-bearing premise

The observed gap between LLM and human choices stems from LLMs accessing pre-existing corpus solutions for canonical games rather than facing equivalent computational bounds.

What would settle it

Running the four proposed tests (conditional dependence, distributional asymmetry, path-dependence under repetition, and paraphrase-robustness) on a canonical game with known corpus solutions and finding that LLM δ matches the structural signatures of human δ would falsify the claim that LLMs bypass the bound.

read the original abstract

Researchers have started using LLM agents in place of human subjects in behavioural and political-science experiments, often as a cheaper substitute for laboratory pools. The substitution does not hold up in strategic settings: humans and LLMs reliably make different choices, and neither fine-tuning on human response data nor persona conditioning has closed the gap. The behavioural-economics literature has, since Simon's introduction of bounded rationality, modelled human strategic behaviour as a classical baseline plus an additive correction term $\delta$. The framework proposed here reads $\delta$ as the mathematical signature of bounded computation: the gap between what an unboundedly-rational agent would compute and what a computationally bounded agent actually produces. For canonical games whose solutions are present in standard training corpora, LLMs retrieve and recombine corpus material, bypassing the bound that produces $\delta$ in humans. The framing extends to reasoning-distilled models through cognitive-hierarchy theory: their accessible level-$k$ strategic reasoning is bounded by compute budget and context length rather than by the cognitive constraints that bound humans, and the $\delta$ they produce, if any, carries different structural signatures. Four operational tests (conditional dependence, distributional asymmetry, path-dependence under repetition, and paraphrase-robustness) are proposed to discriminate human-shaped $\delta$ from LLM-shaped $\delta$. A moderator prediction is that $|\delta|$ scales with peer-signal individuation in the decision environment, with a quantitative bound of Cohen's $d \geq 0.5$ between named-opponent and aggregate-opponent settings.

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 reframes LLM-human gaps in games as LLMs retrieving corpus solutions instead of hitting the same computational bound that creates human δ, but offers only a hypothesis with untested predictions.

read the letter

The main takeaway is that this paper treats the LLM-human mismatch in strategic games as evidence that LLMs bypass bounded computation altogether by pulling answers from training data for games that appear in standard corpora. That distinction from human-style δ is the new piece.

It builds on Simon's bounded-rationality setup and extends it to reasoning-distilled models via cognitive hierarchy, noting that any δ in LLMs would stem from context length and compute budget rather than human cognitive limits. The four proposed tests—conditional dependence on opponent signals, distributional asymmetry, path dependence under repetition, and paraphrase robustness—give concrete ways to check whether observed gaps match retrieval signatures or computation signatures.

The moderator claim that |δ| scales with peer-signal individuation at Cohen's d ≥ 0.5 stands out as unsupported; no derivation or benchmark is given for the threshold. The paper stays conceptual throughout, with no data, no derivations, and no error analysis, so the circularity risk in defining δ as the gap remains unaddressed in practice.

This is aimed at behavioral economists and political scientists already running or planning experiments that substitute LLM agents for human subjects. A reader who needs a structured hypothesis to design follow-up tests will find the operational distinctions useful even if the specific quantitative bound needs work.

The work is coherent enough on its own terms to warrant referee time. The tests are falsifiable in principle, and the framing could help experimenters decide when LLM substitution is likely to distort results.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a bounded-rationality framework in which the correction term δ represents the gap between unbounded-rational and computationally bounded output. It argues that LLMs bypass this δ for canonical games by retrieving and recombining solutions present in training corpora, extends the account to reasoning-distilled models via cognitive-hierarchy considerations, proposes four operational tests (conditional dependence, distributional asymmetry, path-dependence under repetition, and paraphrase-robustness) to discriminate human-shaped from LLM-shaped δ, and advances a moderator prediction that |δ| scales with peer-signal individuation (Cohen's d ≥ 0.5).

Significance. If the proposed tests prove diagnostic, the framework would supply a principled basis for deciding when LLM agents can substitute for human subjects in strategic experiments and would distinguish retrieval signatures from genuine computational bounds. The explicit listing of falsifiable tests is a constructive element of the proposal.

major comments (2)

[Abstract] Abstract: the moderator prediction asserts a quantitative bound of Cohen's d ≥ 0.5 between named-opponent and aggregate-opponent settings without derivation, external benchmark, or supporting calculation; this bound is load-bearing for the claimed empirical testability of the framework.
[Framework (reasoning-distilled models)] Section on the framework extending to reasoning-distilled models: the claim that any δ produced by such models carries different structural signatures because their level-k reasoning is bounded by compute budget and context length (rather than human cognitive constraints) is stated without concrete formal distinction or illustrative example, which underpins the extension of the theory beyond standard LLMs.

minor comments (1)

The four tests are enumerated but their precise operational definitions (e.g., how conditional dependence would be measured in a given game) are not expanded; adding brief implementation sketches would improve clarity without altering the central proposal.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. We address the two major comments point by point below, indicating the revisions we plan to make to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the moderator prediction asserts a quantitative bound of Cohen's d ≥ 0.5 between named-opponent and aggregate-opponent settings without derivation, external benchmark, or supporting calculation; this bound is load-bearing for the claimed empirical testability of the framework.

Authors: We acknowledge that the specific threshold of Cohen's d ≥ 0.5 is stated in the abstract without an accompanying derivation or external benchmark. This value was chosen as a conventional medium effect size in psychological and behavioral research to illustrate the moderator prediction's testability. To address this, we will revise the abstract to clarify that the bound is illustrative and motivated by standard effect-size conventions (e.g., Cohen, 1988), and we will add a brief derivation in the main text linking the moderator prediction to expected differences in peer-signal individuation. This revision will preserve the framework's empirical claims while providing the requested justification. revision: yes
Referee: [Framework (reasoning-distilled models)] Section on the framework extending to reasoning-distilled models: the claim that any δ produced by such models carries different structural signatures because their level-k reasoning is bounded by compute budget and context length (rather than human cognitive constraints) is stated without concrete formal distinction or illustrative example, which underpins the extension of the theory beyond standard LLMs.

Authors: We agree that the extension to reasoning-distilled models requires a more explicit formal distinction and example to be fully convincing. In the revised version, we will introduce a short subsection or paragraph providing a concrete example using the p-beauty contest game. We will contrast how a standard LLM might retrieve the Nash equilibrium (bypassing δ) with how a reasoning-distilled model, limited by context length to lower k levels, produces a δ with different properties (e.g., sensitivity to compute budget rather than cognitive load). This will include a formal note distinguishing the bounding mechanisms: human cognitive constraints vs. model-specific resource limits. We believe this addition will solidify the theoretical extension. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The manuscript is a conceptual proposal that reinterprets the existing bounded-rationality term δ (defined in the literature as the gap between unbounded-rational baseline and observed behavior) as a computational signature and contrasts it with LLM corpus-retrieval behavior. It advances four operational tests and one moderator prediction as hypotheses to discriminate signatures, without any internal equations, fitted parameters, or self-citations that reduce a claimed result to its own inputs by construction. The argument is therefore self-contained and relies on external empirical discrimination rather than definitional closure.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that human strategic behavior follows a rational baseline plus additive δ from bounded computation, and on the interpretation that LLMs bypass this via corpus access; the Cohen's d bound appears as an asserted quantitative prediction without independent grounding.

free parameters (1)

Cohen's d >= 0.5 moderator bound
The quantitative threshold for the predicted scaling of |δ| with peer-signal individuation is stated without derivation or fitting procedure shown.

axioms (1)

domain assumption Human strategic behavior is a classical baseline plus additive correction δ arising from bounded computation
Invoked at the start of the framework as the modeling approach since Simon.

pith-pipeline@v0.9.1-grok · 5808 in / 1424 out tokens · 58508 ms · 2026-07-01T16:53:15.482998+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

I., & Kalai, A

Aher, G., Arriaga, R. I., & Kalai, A. T. (2023). Using large language models to simulate multiple humans and replicate human subject studies. In Proceedings of the 40th International Conference on Machine Learning (ICML). arXiv:2208.10264. https://arxiv.org/abs/2208.10264 Akata, E., Schulz, L., Coda-Forno, J., Oh, S. J., Bethge, M., & Schulz, E. (2025). P...

work page doi:10.1017/pan.2023.2 2023
[2]

https://doi.org/10.1016/j.tics.2024.01.011 Mei, Q., Xie, Y., Yuan, W., & Jackson, M. O. (2024). A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences, 121(9), Article e2313925121. https://doi.org/10.1073/pnas.2313925121 Mozikov, M., Severin, N., Bodishtianu, V., Glushanina, M., Baklashkin,...

work page doi:10.1016/j.tics.2024.01.011 2024

[1] [1]

I., & Kalai, A

Aher, G., Arriaga, R. I., & Kalai, A. T. (2023). Using large language models to simulate multiple humans and replicate human subject studies. In Proceedings of the 40th International Conference on Machine Learning (ICML). arXiv:2208.10264. https://arxiv.org/abs/2208.10264 Akata, E., Schulz, L., Coda-Forno, J., Oh, S. J., Bethge, M., & Schulz, E. (2025). P...

work page doi:10.1017/pan.2023.2 2023

[2] [2]

https://doi.org/10.1016/j.tics.2024.01.011 Mei, Q., Xie, Y., Yuan, W., & Jackson, M. O. (2024). A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences, 121(9), Article e2313925121. https://doi.org/10.1073/pnas.2313925121 Mozikov, M., Severin, N., Bodishtianu, V., Glushanina, M., Baklashkin,...

work page doi:10.1016/j.tics.2024.01.011 2024