Pith. sign in

REVIEW 3 major objections 7 minor 28 references

Frontier language models show stable, model-specific risk styles in poker, and they adapt differently when stakes or their own chips change.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 13:10 UTC pith:WXK7XMG7

load-bearing objection Clean multi-model poker assay shows real, stable VPIP/PFR risk profiles and heterogeneous adaptation across six frontier LLMs; external validity is the only serious soft spot. the 3 major comments →

arxiv 2607.10251 v1 pith:WXK7XMG7 submitted 2026-07-11 cs.AI

Behavioural Signatures of Risk-Sensitive Decision-Making in Large Language Models

Xuankun Rong , Wenke Huang , Bo Du , Dacheng Tao , Mang Ye This is my paper
classification cs.AI
keywords large language modelsrisk-sensitive decision-makingbehavioural profilesTexas Hold'emParticipation VPIPProactiveness PFRmulti-agent interactionrisk adaptation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether large language models used for decisions under uncertainty have stable, readable risk habits rather than acting as interchangeable engines. Using six-player no-limit Texas Hold'em as a controlled testbed, the authors measure two behaviours: Participation (how often a model voluntarily enters a hand) and Proactiveness (how often it raises pre-flop). Across many 100-hand blocks, six frontier models form a clear spectrum from conservative to aggressive. Those profiles mostly hold when models face each other instead of copies of themselves, though the most conservative and most aggressive models pull further apart. When blinds rise for everyone or a single model's stack shrinks, models do not all tighten the same way: some contract broadly, some mainly stop escalating, and one stays nearly fixed. The practical point is that risk style and risk plasticity are measurable properties that can be audited before models are trusted in interactive decision settings.

Core claim

Frontier LLMs express stable, model-specific risk profiles in the Participation–Proactiveness plane, forming a conservative-to-aggressive spectrum that largely survives mixed-opponent play while extremes diverge further; under rising global stakes or personal short stacks they adapt in structured but heterogeneous ways, from broad contraction through selective de-escalation to near-invariance, showing they differ both in baseline disposition and in which risk signals they heed.

What carries the argument

The two-metric behavioural state space of Participation (VPIP: voluntary pre-flop entry rate) and Proactiveness (PFR: pre-flop raise rate), measured over independent 100-hand blocks in homogeneous self-play versus heterogeneous mixed tables and under global blind pressure versus personal short-stack constraints.

Load-bearing premise

The central claim rests on the premise that pre-flop entry and raise rates in a fixed-prompt, six-player Hold'em simulation are a fair enough proxy for general risk-sensitive decision-making in real interactive systems.

What would settle it

Run the same models on a non-poker sequential risk task (for example multi-round resource allocation or medical triage under incomplete information) with analogous voluntary-entry and escalation metrics; if model orderings and adaptation patterns scramble or vanish, the claim that these are general risk signatures fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript studies whether frontier LLMs exhibit stable, interpretable risk-sensitive decision regularities using a controlled six-player no-limit Texas Hold’em assay. Behaviour is summarized by Participation (VPIP) and Proactiveness (PFR). In homogeneous self-play, six models occupy a conservative-to-aggressive spectrum; most profiles remain robust in heterogeneous mixed tables, while the most conservative and aggressive models diverge further. Outcome tables show that these styles redistribute rewards under mixed competition. Under rising global blind pressure and under personal short-stack constraints, models adapt heterogeneously (broad contraction, selective de-escalation, near-invariance). The authors frame this as a two-component (trait + adaptive) behavioural basis for auditing risk-sensitive LLM decision systems, with public code and logs.

Significance. If the within-assay regularities hold, the paper supplies a concrete, reproducible behavioural audit layer for interactive LLM decision systems that goes beyond accuracy benchmarks. Strengths include a clean homogeneous-vs-heterogeneous design, two orthogonal risk manipulations, block-level statistics, standard external poker metrics rather than ad-hoc scores, and full public code plus simulation logs. The descriptive spectrum and plasticity patterns are useful for model selection and multi-agent deployment even if they do not establish human-like risk psychology. The contribution is primarily empirical and methodological rather than theoretical, but it is timely for AI evaluation and multi-agent systems venues.

major comments (3)
  1. [Methods] Methods (Experimental setup / Models evaluated): decoding and sampling hyperparameters (temperature, top-p/top-k, seed policy, max tokens, provider defaults) are not reported. For a central claim of stable model-specific behavioural profiles, these settings are load-bearing: nonzero temperature or provider-default stochasticity can inflate within-model variance and alter VPIP/PFR. Please report exact API parameters for every model and, if temperature > 0, quantify sensitivity of the spectrum and Mann–Whitney results to temperature (or re-run main conditions at temperature 0).
  2. [Results / Personal risk exposure] Results (Personal risk exposure; Fig. 4): personal short-stack conditions use only 10 independent runs versus 30 for homogeneous/heterogeneous and global-risk conditions. The claim that models become more cautious under personal resource constraint, with model-specific forms of caution, is therefore much more weakly supported than the trait and global-pressure claims. Either increase n to match the main protocol or clearly demote personal-risk findings to exploratory and avoid parallel strength of language with the 30-block results.
  3. [Results / Statistical analysis] Results (Social robustness; Fig. 2) and Statistical analysis: significance of mixed-table shifts is assessed with two-sided Mann–Whitney U tests at p<0.05 per model and metric, without reported effect sizes, confidence intervals, or multiple-comparison control across six models × two metrics. The load-bearing claim that extremes diverge while mid-spectrum models are robust needs effect sizes (e.g., rank-biserial or median differences with CIs) and either pre-specified primary contrasts or multiplicity-aware reporting so that “significant” extreme shifts are not over-read.
minor comments (7)
  1. [Abstract / Discussion] Abstract and Discussion somewhat over-extend from a single fixed-prompt poker assay to “auditing risk-sensitive decision-making in interactive settings.” The Discussion already notes the proxy nature of poker; tighten abstract wording so the claim is scoped to incomplete-information sequential betting under this protocol unless non-poker validation is added.
  2. [Table 2] Table 2 risk-adaptation slopes are linear fits over four discrete blind levels; report R² or residual diagnostics, and clarify whether slopes are fit on block means or pooled rates. Near-zero Gemini slopes should be distinguished from “insensitive” versus “already floor-compressed” behaviour.
  3. [Figure 2] Figure 2 caption in the body text inconsistently names “GPT-5.4 and Gemini-3.1-Pro” vs “GPT and Gemini”; align naming with Table 3 identifiers throughout.
  4. [Methods / A.3] Methods: state whether models were allowed multi-turn tool use or only single-shot JSON actions, and report retry rates by model (Appendix A shows retries up to 55). Output-format reliability should be summarized as a secondary metric so behavioural rates are not confounded with parser recovery.
  5. [Results] No human or strong poker-bot baseline under the same prompt/state schema is provided. Even a brief reference policy (e.g., random legal action or a simple heuristic) would calibrate absolute VPIP/PFR levels; this is optional but would strengthen interpretability.
  6. [Results] Minor duplication: the Results opening and the paragraph before Results repeat nearly identical text about profiles remaining robust and extremes diverging; compress for length.
  7. [References / Table 3] References include system cards and technical reports with future-dated links; ensure citation metadata matches the exact model snapshots used in Tab. 3.

Circularity Check

0 steps flagged

No circularity: observational poker metrics and descriptive slopes do not force the claimed risk profiles or adaptation patterns by construction.

full rationale

The paper is an empirical multi-model assay, not a first-principles derivation. Participation (VPIP) and Proactiveness (PFR) are standard external poker definitions applied to logged pre-flop actions (Eqs. 1–2); they are not defined in terms of the spectrum or plasticity claims. Homogeneous vs heterogeneous separations, Mann–Whitney tests, outcome tables, and risk-adaptation slopes (linear fits of observed rates across four blind levels; shift magnitude as Euclidean norm of those slopes) are descriptive summaries of measured block-level rates, not fitted parameters that redefine or statistically force the target regularities. Citations are to classic decision literature, prior poker AI, and model cards; none is a load-bearing uniqueness theorem or ansatz from the present authors that closes the argument. The within-assay claims are therefore self-contained against the public logs and methods; external-validity limits of the poker proxy are stated but do not constitute circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard poker metric definitions, the experimental operationalization of risk (blinds and stacks), and the domain premise that this closed simulation is informative about risk-sensitive LLM decision support. No free parameters are fitted to produce the spectrum; stack/blind levels and session counts are design choices. No new physical or mathematical entities are postulated.

free parameters (2)
  • blind/stack risk levels (10/50/100/200 BB; stacks 1000/500/100) = design levels only
    Chosen by experimenters to operationalize global and personal risk; not fitted to maximize separation but still discrete design knobs that define the adaptation axes.
  • session length and count (100-hand blocks; 30 or 10 runs) = 30 main / 10 short-stack
    Statistical unit size chosen for block-level inference; affects variance estimates but not the metric definitions themselves.
axioms (4)
  • domain assumption VPIP and PFR, normalized by all hands, are valid primary measures of risk disposition and risk escalation for LLM agents.
    Stated in Methods/Behavioural metrics; borrowed from poker analytics without independent validation that they capture non-poker decision risk.
  • domain assumption A fixed research-simulation prompt and identical state schema make cross-model behavioural differences attributable to model policy rather than instruction wording.
    Methods Experimental setup; necessary for comparability but untested against prompt variants.
  • domain assumption No-limit Texas Hold'em with incomplete information and sequential betting is a suitable controlled proxy for real-world risk-sensitive decision support.
    Introduction and Discussion; load-bearing for the auditing claim.
  • standard math Block-level Mann–Whitney tests on independent 100-hand sessions adequately support claims of stability and significant shifts.
    Statistical analysis section; standard non-parametric comparison.
invented entities (1)
  • model-specific risk plasticity / adaptation patterns (broad contraction, selective de-escalation, near-invariant) no independent evidence
    purpose: Label heterogeneous responses to risk pressure beyond baseline trait scores.
    Descriptive taxonomy derived from observed slopes; not an independent latent variable with external measurement.

pith-pipeline@v1.1.0-grok45 · 33865 in / 2771 out tokens · 31026 ms · 2026-07-14T13:10:35.427557+00:00 · methodology

0 comments
read the original abstract

As large language models (LLMs) are increasingly used in decision support, it is important to understand whether their choices under uncertainty exhibit stable and interpretable behavioural regularities. Human decision-making combines relatively persistent risk preferences with context-dependent adjustment, yet it remains unclear whether analogous behavioural structure can be observed in LLM-based decision systems. Here we examine this question using a controlled multi-model framework based on no-limit Texas Hold'em, where behaviour is quantified by Participation, measuring voluntary engagement in uncertain opportunities, and Proactiveness, measuring pre-flop risk escalation. Across homogeneous self-play and heterogeneous mixed-model interactions, frontier LLMs exhibit stable, model-specific risk profiles, forming a spectrum from conservative to aggressive decision styles. These profiles remain largely robust under changing opponent composition, while the most conservative and most aggressive models diverge further in mixed settings. Under global risk pressure and personal resource constraint, models adapt in structured but heterogeneous ways, ranging from broad behavioural contraction to selective de-escalation and near-invariant behaviour. These findings suggest that LLMs differ not only in baseline risk disposition, but also in the risk signals they respond to and the flexibility with which they adjust, providing a behavioural basis for auditing risk-sensitive decision-making in interactive settings. Our code is publicly available at: https://github.com/XuankunRong/AgentTexasPoker.

Figures

Figures reproduced from arXiv: 2607.10251 by Bo Du, Dacheng Tao, Mang Ye, Wenke Huang, Xuankun Rong.

Figure 1
Figure 1. Figure 1: Framework for evaluating risk-sensitive decision-making in large language models. a. Decision-making under uncertainty increasingly involves AI systems in risk-sensitive contexts, motivating the question of whether LLMs exhibit stable and adaptive behavioural signatures of risk. b. We use interactive Texas Hold’em as a controlled behavioural assay, quantifying decisions with Participation (VPIP), which mea… view at source ↗
Figure 2
Figure 2. Figure 2: Stable trait-level risk profiles across homogeneous and heterogeneous interactions. a. Experimental settings for homogeneous self-play, where all six table positions use the same model, and heterogeneous mixed play, where six different models interact at the same table. b. Participation rates across the two settings. Models show clear separation and limited within-model variation. c. Proactiveness rates ac… view at source ↗
Figure 3
Figure 3. Figure 3: Behavioural adaptation under increasing global risk pressure. Participation and Proactiveness rates are shown for each model across four risk levels, corresponding to big blinds of 10, 50, 100 and 200 chips with a fixed initial stack of 1,000 chips. Increasing risk pressure generally reduces both engagement and aggression, but models follow distinct trajectories: broad contraction in Claude, Xiaomi and Dee… view at source ↗
Figure 4
Figure 4. Figure 4: Behavioural adaptation under personal risk exposure. Participation and Proactiveness rates are shown for each model under the full-stack baseline and two short-stack conditions, where the focal model’s stack is reduced to 500 or 100 chips while the surrounding environment is kept fixed. Personal risk exposure generally reduces engagement and aggression, but models follow distinct adaptation patterns: Xiaom… view at source ↗

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

28 extracted references · 5 linked inside Pith

  1. [1]

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

    2011

  2. [2]

    E., Bloemen, P

    Marchau, V ., Walker, W. E., Bloemen, P. & Popper, S. W. Decision making under deep uncertainty: from theory to practice(Springer, 2019)

    2019

  3. [3]

    Achiam, J.et al.Gpt-4 technical report.arXiv preprint arXiv:2303.08774(2023)

    Pith/arXiv arXiv 2023

  4. [4]

    InInternational Conference on Learn- ing Representations(2023)

    Yao, S.et al.React: Synergizing reasoning and acting in language models. InInternational Conference on Learn- ing Representations(2023)

    2023

  5. [5]

    neural information process- ing systems36, 68539–68551 (2023)

    Schick, T.et al.Toolformer: Language models can teach themselves to use tools.Adv. neural information process- ing systems36, 68539–68551 (2023)

    2023

  6. [6]

    S.et al.Generative agents: Interactive simulacra of human behavior

    Park, J. S.et al.Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technol- ogy, 1–22 (2023)

    2023

  7. [7]

    & Morgenstern, O.Theory of Games and Economic Behavior(Princeton University Press, Princeton, NJ, 1944)

    von Neumann, J. & Morgenstern, O.Theory of Games and Economic Behavior(Princeton University Press, Princeton, NJ, 1944)

    1944

  8. [8]

    J.The foundations of statistics(Courier Cor- poration, 2012)

    Savage, L. J.The foundations of statistics(Courier Cor- poration, 2012)

    2012

  9. [9]

    & Tversky, A

    Kahneman, D. & Tversky, A. Prospect theory: An analy- sis of decision under risk. InHandbook of the fundamen- tals of financial decision making: Part I, 99–127 (World Scientific, 2013)

    2013

  10. [10]

    & Kahneman, D

    Tversky, A. & Kahneman, D. Advances in prospect theory: Cumulative representation of uncertainty.J. Risk uncertainty5, 297–323 (1992)

    1992

  11. [11]

    U., Blais, A.-R

    Weber, E. U., Blais, A.-R. & Betz, N. E. A domain- specific risk-attitude scale: Measuring risk perceptions and risk behaviors.J. behavioral decision making15, 263–290 (2002)

    2002

  12. [12]

    & Weber, E

    Blais, A.-R. & Weber, E. U. A domain-specific risk- taking (dospert) scale for adultpopulations.Judgm. Decis. making1, 33–47 (2006)

    2006

  13. [13]

    Schoemaker, P. J. Determinants of risk-taking: Behav- ioral and economic views.J. Risk Uncertain.6, 49–73 (1993)

    1993

  14. [14]

    Wang, L.et al.A survey on large language model based autonomous agents.Front. Comput. Sci.18, 186345 (2024)

    2024

  15. [15]

    & Roesler, E

    Fahnenstich, H., Rieger, T. & Roesler, E. Trusting un- der risk–comparing human to ai decision support agents. Comput. Hum. Behav.153, 108107 (2024)

    2024

  16. [16]

    & Kumar, A

    Steyvers, M. & Kumar, A. Three challenges for ai- assisted decision-making.Perspectives on Psychol. Sci. 19, 722–734 (2024)

    2024

  17. [17]

    W., Bettman, J

    Payne, J. W., Bettman, J. R. & Johnson, E. J.The adaptive decision maker(Cambridge university press, 1993)

    1993

  18. [18]

    Sitkin, S. B. & Pablo, A. L. Reconceptualizing the deter- minants of risk behavior.Acad. management review17, 9–38 (1992)

    1992

  19. [19]

    Palomäki, J., Laakasuo, M., Cowley, B. U. & Lappi, O. Poker as a domain of expertise.J. expertise3, 66–87 (2020)

    2020

  20. [20]

    & Sandholm, T

    Brown, N. & Sandholm, T. Superhuman ai for heads-up no-limit poker: Libratus beats top professionals.Science 359, 418–424 (2018)

    2018

  21. [21]

    & Sandholm, T

    Brown, N. & Sandholm, T. Superhuman ai for multiplayer poker.Science365, 885–890 (2019)

    2019

  22. [22]

    Moravˇcík, M.et al.Deepstack: Expert-level artificial intelligence in heads-up no-limit poker.Science356, 508–513 (2017)

    2017

  23. [23]

    Singh, A.et al.Openai gpt-5 system card.arXiv preprint arXiv:2601.03267(2025)

    Pith/arXiv arXiv 2025

  24. [24]

    Anthropic system cards

    Anthropic. Anthropic system cards. https://www. anthropic.com/system-cards (2026)

    2026

  25. [25]

    Team, G.et al.Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  26. [26]

    Guo, D.et al.Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.Nature645, 633–638 (2025)

    2025

  27. [27]

    arXiv preprint arXiv:2505.07608(2025)

    Xiaomi, L.et al.Mimo: Unlocking the reasoning poten- tial of language model–from pretraining to posttraining. arXiv preprint arXiv:2505.07608(2025)

    Pith/arXiv arXiv 2025

  28. [28]

    action":

    Yang, A.et al.Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025). 11/27 A Supplementary Prompt and Decision Case Records This appendix gives the prompt contract, state representation, output parser and representative decision case records used to audit the analyses in the main text. The cases are not additional statistical evidence by themselves...

    Pith/arXiv arXiv 2025