Pith. sign in

REVIEW 4 major objections 5 minor 9 references

Strategy Adaptation in Large Language Model Werewolf Agents

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Switching between Support and Attack prompts, chosen from role estimates, raises Werewolf agents' win rates in the Werewolf game.

desk verdict A promising pilot on explicit strategy adaptation for LLM Werewolf agents, but 30-game win rates without error bars cannot support the headline claim. read the letter →

arxiv 2507.12732 v1 pith:SZXLDBYT submitted 2025-07-17 cs.CL

classification cs.CL
keywords WerewolfgameLLMagentsstrategyadaptationroleestimationpromptengineeringsocialdeductionSupport/Attackstrategieswin-rateanalysis
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

The paper aims to show that explicit strategy adaptation improves large-language-model agents in the Werewolf game: instead of following one fixed prompt strategy, an agent chooses between a Support strategy (agreeing with and defending likely teammates) and an Attack strategy (casting doubt on likely opponents), using conversation context plus explicit role estimates. The claim matters because prior prompt-engineering agents for Werewolf embed a single strategy and cannot adjust when suspicion shifts, while human play requires flexible switching. In experiments with GPT-4o-mini and Gemini 2.0 Flash, the adaptive agents reached the highest win rates for the Werewolf side in both models, and for the Villager side with GPT, though not with Gemini. An ablation shows role estimation drives the Werewolf gain but appears to hurt Villager agents, whose win rate fell from 0.80 to 0.73 when estimation was added.

What carries the argument

The mechanism is a two-strategy prompt library plus an adaptation gate. Each agent first estimates every other player's role on a 0-4 scale for werewolf, villager, seer, and doctor, after stating its reasoning; then, at three moments per round (after night actions, after the day debate, and after voting), the LLM reads the conversation history and these role scores and picks either the Support or Attack prompt according to criteria such as 'use Support when you are being suspected' and 'use Attack when a likely werewolf is already under suspicion'. The chosen prompt then generates the agent's utterances for that phase, so the strategy is part of the context the model responds to, not a separate controller.

What would settle it

Run the same 30-match protocol with the true roles substituted for the LLM's role-estimation scores; if Villager win rates still fall below the 0.80 no-estimation baseline, the adaptation criteria themselves are the bottleneck, not the estimator. If they rise above it, the estimator is the bottleneck and better role-estimation prompts should transfer directly to higher win rates.

Watch

Extended reading notes

Core claim

The central claim is that dynamically switching between two predefined strategy prompts—Support and Attack—guided by a hand-written decision rule and by each agent's 0-4 role estimates for other players, raises the win rate of LLM Werewolf agents relative to fixed-strategy and implicit-strategy baselines. The authors find the effect is most consistent for the Werewolf side: adaptation produced the highest Werewolf win rate with both tested models, and the Attack strategy chosen under adaptation was followed by a smaller increase in how accurately others estimated the werewolves' roles, suggesting it helps them act without arousing suspicion. For Villager agents the result is mixed; GPT Villagers improved under adaptation, Gemini Villagers did not, and the ablation attributes part of the difficulty to low role-estimation accuracy rather than to the adaptation rule itself.

Load-bearing premise

The method assumes the LLM's 0-4 role estimates are accurate enough to choose between Support and Attack; the paper's own ablation shows Villagers win less with estimation (0.73 vs 0.80) and werewolves are estimated at only about 0.30, so this premise is not clearly satisfied.

Editorial extensions

If this is right

  • Werewolf agents in other social-deduction games could adopt the same Support/Attack switch and expect the largest gains, since they know their teammates and can coordinate around the choice.
  • The value of adaptation depends on role-estimation quality; improving the estimator should improve Villager performance, where the paper currently reports a drop from 0.80 to 0.73 when estimation is added.
  • The three switching points (night, post-debate, post-vote) give a natural schedule for learning when to switch; the paper suggests reinforcement learning as a future avenue.
  • Because the Attack strategy lowered how much the werewolves' roles were detected, adaptive deception can be measured directly through changes in role-estimation accuracy rather than only through final win rates.

Reading between the lines

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

  • If role estimates were replaced with ground-truth roles, the Villager side would likely match or beat the no-estimation baseline; the paper's 0.30 estimate for werewolves suggests the current prompt is too weak to support Villager decisions.
  • The adaptation rule may be overfit to the Werewolf side: the criteria were designed around suspicion and credibility, which are exactly the levers werewolves control, while villagers need a third option (revealing information as seer or doctor) that the Support/Attack dichotomy does not provide.
  • A direct test of the method would randomize the order of the three adaptation moments or run adaptation only at one moment, isolating whether switching frequency or rule content drives the win-rate gain.
  • The same two-prompt adaptation could be ported to other negotiation settings where agents must choose between cooperating with allies and attacking opponents, with role estimation replaced by trust estimation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a strategy-adaptation mechanism for LLM-based Werewolf agents, in which agents switch between a Support strategy and an Attack strategy based on conversation context and explicit role estimates. The method is evaluated in the Werewolf Arena environment using GPT-4o-mini and Gemini-2.0-flash, comparing the adaptive agent against an implicit-strategy baseline and fixed Support/Attack prompts, plus an ablation that removes adaptation and/or role estimation. The paper reports win rates over 30 matches per condition and concludes that adaptation 'consistently outperforms' fixed strategies and prior methods for Werewolf agents, while acknowledging that Villager agents sometimes do worse with adaptation, especially for Gemini and in the role-estimation ablation.

Significance. If the claimed effect is real, the paper would provide a simple, prompt-only method for dynamic strategy selection in social-deduction games, a task that demands situational flexibility well beyond static prompting. The authors are transparent about their limitations, provide the full prompts in an appendix, and include an ablation study. However, the empirical support is currently too thin to establish the central claim: win rates are computed from only 30 games per condition without confidence intervals or significance tests, and the paper's own ablation shows that role estimation can hurt Villager performance. The conclusion's reference to outperforming 'prior methods' is also unsupported because no prior method is implemented in the experiments. With improved statistical analysis and more modest claims the work could be a useful contribution, but in its present form the evidence is not convincing.

major comments (4)
  1. [Section 6, Conclusion] The conclusion states that the proposal is 'consistently outperforming both fixed strategies and prior methods in win rates,' but the experiments in Section 4 only compare against the implicit Werewolf Arena baseline and the two fixed-strategy prompts (Support and Attack). No experiments against Kano et al. (2023), Xu et al. (2024a), or Wu et al. (2024) are reported, so the claim about prior methods is not supported by the presented data.
  2. [Section 5.1, Figure 2, and Table 2] The headline comparisons rely entirely on 30 games per condition with no error bars, confidence intervals, or significance tests. For the key Werewolf ablation result in Table 2, a win rate of 0.60 versus 0.40 corresponds to a two-sided Fisher exact p-value of roughly 0.18, so the observed 0.20 gap could easily arise from sampling noise; the 0.08 gap against the 0.52 condition is even weaker. The phrase 'consistently outperforming' is therefore not established by the reported statistics.
  3. [Section 5.2, Table 2] The ablation results are mixed and partly undermine the claim that role estimation helps. For Villagers, removing role estimation improves the win rate from 0.73 to 0.80, which the paper explains by low estimation accuracy (Est≈0.30). However, no statistical test is provided to show that the 0.73/0.80 difference is meaningful, and the text in Section 5.2 says role estimation 'had a significant impact' without a significance test to support that wording.
  4. [Section 5.3 and Appendix A.2] The role-estimation accuracy for Werewolves is reported only as Est≈0.30, which is barely above the chance level of 0.25 for four roles. Since strategy selection in the adaptation method relies on these role estimates (per the criteria in Appendix A.2), the authors should report the complete Est values and formally compare them against the chance baseline to justify that the estimates provide usable signal; as presented, the near-chance accuracy casts doubt on the proposed mechanism.
minor comments (5)
  1. [Section 1] The sentence 'Success in the game relies on The game requires both the ability of deductive reasoning...' contains a duplicated and garbled phrase; it should be corrected to a single coherent sentence.
  2. [Appendix A.4, Tables 7, 8, and 9] Table 7 and Table 9 headers misspell 'Attack' as 'Atack,' and Table 8's header misspells 'Werewolves' as 'Worewolves.'
  3. [Appendix A.4, Table 5] The role-estimation prompt includes the text 'num villagers' without substitution; since the game has four villagers, the prompt should state the actual number rather than leaving a placeholder.
  4. [Section 5.3, Table 3] The table reports average changes in Est but does not include standard deviations or any measure of variance, so it is difficult to assess whether the observed differences between adaptation and fixed strategies are meaningful.
  5. [Section 7, Limitations] The limitations section is honest about the small number of models and the lack of human evaluation, but it would be helpful to also state whether code and prompt templates will be released for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: strategy adaptation is empirically evaluated against held-out 30-match win rates with no fitted parameter or self-citation chain forcing the outcome.

full rationale

The paper's load-bearing claim is that dynamically selecting between hand-crafted Support and Attack strategy prompts improves Werewolf-side win rates relative to fixed or implicit strategies. This claim is supported entirely by measured 30-match win rates (Figure 2 and Table 2) comparing the proposal, baselines, and ablations. No parameter is fitted to the reported outcomes and then renamed as a prediction; the strategy-selection criteria in Appendix A.2 are fixed, hand-written prompts, and the role-estimation scores are measured quantities (Equations 1 and 2) reported as analysis rather than used to mathematically derive the win rates. The ablation subtracts components one at a time, so the comparisons are direct empirical contrasts rather than identities by construction. There is no load-bearing self-citation: the prior work cited (Werewolf Arena, AIWolfDial, and earlier Werewolf-agent papers) is used as environment and baseline, not as an unverified premise that forces the conclusion. The admitted limitation that Villager performance sometimes decreases with adaptation and the absence of significance testing are statistical-strength concerns, not circularity concerns. Thus the derivation chain is self-contained as an empirical evaluation, and no circular step is present.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The method introduces no fitted free parameters. It relies on the Werewolf Arena rules as an external environment, on the LLM's ability to follow strategy and role-estimation prompts, and on the assumption that 30 matches are enough to distinguish win rates. The role estimation accuracy is an internal component, not a fitted constant, but its reliability is a key empirical assumption.

assumptions (4)
  • domain assumption Werewolf Arena rules and configuration are an appropriate testbed.
    The paper adopts the 8-player role distribution from Bailis et al. (2024) without validating generalization to other settings.
  • domain assumption LLM agents execute the Support/Attack prompts and the adaptation criteria as intended.
    The paper only shows two utterance examples (A.1); no systematic compliance check is reported.
  • domain assumption Role estimation scores (0 to 4) reliably indicate role likelihood.
    Strategy selection uses these scores; the paper reports Est approximately 0.30 for werewolves, suggesting low accuracy.
  • domain assumption Win rates over 30 matches provide statistically meaningful comparison.
    No confidence intervals or significance tests are given, and 30 games have large variance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Strategy Adaptation in Large Language Model Werewolf Agents." pith.science (2026). https://pith.science/paper/SZXLDBYT

@misc{pith2026250712732,
  author       = {Pith},
  title        = {Pith review of: Strategy Adaptation in Large Language Model Werewolf Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SZXLDBYT}},
  note         = {Machine review of arXiv:2507.12732}
}
read the original abstract

This study proposes a method to improve the performance of Werewolf agents by switching between predefined strategies based on the attitudes of other players and the context of conversations. While prior works of Werewolf agents using prompt engineering have employed methods where effective strategies are implicitly defined, they cannot adapt to changing situations. In this research, we propose a method that explicitly selects an appropriate strategy based on the game context and the estimated roles of other players. We compare the strategy adaptation Werewolf agents with baseline agents using implicit or fixed strategies and verify the effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2507.12732 by the authors.

Figure 1
Figure 1. Overview of strategy adaptation. Role Num Side Ability villager 4 Villager None seer 1 Villager Reveal one’s identity doctor 1 Villager Save one from attack werewolf 2 Werewolves Kill one [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Win rates of agents under different strategy [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

9 extracted references · 2 canonical work pages

  1. [1]

    Suma Bailis, Jane Friedhoff, and Feiyang Chen. 2024. https://arxiv.org/abs/2407.13943 Werewolf Arena: A Case Study in LLM Evaluation via Social Deduction . Preprint, arXiv:2407.13943

  2. [2]

    Yoshinobu Kano, Neo Watanabe, Kaito Kagaminuma, Claus Aranha, Jaewon Lee, Benedek Hauer, Hisaichi Shibata, Soichiro Miki, Yuta Nakamura, Takuya Okubo, Soga Shigemura, Rei Ito, Kazuki Takashima, Tomoki Fukuda, Masahiro Wakutani, Tomoya Hatanaka, Mami Uchida, Mikio Abe, Akihiro Mikami, Takashi Otsuki, Zhiyang Qi, Kei Harada, Michimasa Inaba, Daisuke Katagam...

  3. [3]

    Xuehong Tao, Zhiqi Shen, and Chunyan Miao. 2010. https://api.semanticscholar.org/CorpusID:14330304 A Cooperative-Competitive Negotiation Model

  4. [4]

    Tomer Ullman. 2023. https://arxiv.org/abs/2302.08399 Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks . Preprint, arXiv:2302.08399

  5. [5]

    Shuang Wu, Liwen Zhu, Tao Yang, Shiwei Xu, Qiang Fu, Yang Wei, and Haobo Fu. 2024. https://arxiv.org/abs/2402.02330 Enhance Reasoning for Large Language Models in the Game Werewolf . Preprint, arXiv:2402.02330

  6. [6]

    Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. 2024 a . https://arxiv.org/abs/2309.04658 Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf . Preprint, arXiv:2309.04658

  7. [7]

    Zelai Xu, Chao Yu, Fei Fang, Yu Wang, and Yi Wu. 2024 b . https://arxiv.org/abs/2310.18940 Language Agents with Reinforcement Learning for Strategic Play in the Werewolf Game . Preprint, arXiv:2310.18940

  8. [8]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

Show all 9 references
  1. [9]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.