Pith. sign in

REVIEW 3 major objections 7 minor 30 references

Persistent public suspicion scores turn the Mafia game into a history-dependent race to two absorbing ends, with a simple early-time law and a finite-size collapse when no detectives are present.

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 ↗

History-dependent suspicion scores in an agent Mafia model yield F(τ)∼(τ/N)^{N_m} early extinction without detectives and an empirical N_c collapse of win probabilities that detectives break.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Real intermediate mechanism and honest dilute-limit check; the N_c collapse is useful but thinner than the abstract sells, and the phase-transition language is overreach. the 3 major comments →

arxiv 2607.24980 v1 pith:CMOUSGJB submitted 2026-07-27 physics.soc-ph cond-mat.stat-mech

Cumulative suspicion and absorption dynamics in an agent-based Mafia game

Eduardo Velasco Stock , Roberto da Silva This is my paper
classification physics.soc-ph cond-mat.stat-mech
keywords Mafia gameWerewolfagent-based modelcumulative suspicionabsorbing statesfinite-size scalinghidden-role gamesMonte Carlo
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.

The reading

This paper replaces the usual random-vote day phase of the Mafia (Werewolf) game with cumulative, player-specific suspicion scores updated by role-dependent pairwise accusations. Those scores make each public elimination history-dependent, so two games with the same faction sizes can still have different kill probabilities. Without detectives, a random-execution approximation says that early on, when the mafia is dilute, the chance the mafia is already wiped out by time τ scales as (τ/N) to the power of the initial mafia count—and simulations match that as the mafia fraction shrinks. Winning-probability curves for different starting mafia sizes also collapse onto one master curve once population is rescaled by an empirical crossover size Nc where the good side’s win chance is one half. Adding perfectly informed detectives shortens mafia survival and breaks that one-parameter collapse, showing that the internal makeup of the good faction matters. The model is offered as a minimal bridge from microscopic accusation histories to the macroscopic absorption statistics of hidden-role contests.

Core claim

In an agent-based Mafia game driven by persistent public suspicion rather than uniform random voting, the early-time cumulative probability of mafia extinction in the dilute regime follows F(τ)∼(τ/N)^{N_m} without detectives, and good-faction winning probabilities for different initial mafia sizes approximately collapse when total population is rescaled by a single crossover scale Nc; perfectly informed detectives both speed extinction and destroy that simple collapse.

What carries the argument

Cumulative public suspicion scores: randomly paired agents add role-dependent increments to player-specific scores that persist across rounds, and the day elimination is drawn with probability proportional to each survivor’s accumulated score, yielding a history-dependent path to one of two absorbing outcomes (mafia extinct or mafia–good parity).

Load-bearing premise

The day-phase accusation rules—who smears whom in each pair type, including perfect detective hits and half-weight unpaired agents—are stipulated by hand rather than taken from real play or equilibrium strategy; if those rules are wrong, the power law and the Nc collapse need not hold.

What would settle it

Run large Monte Carlo ensembles with dilute mafia and no detectives and check whether the early-time cumulative extinction probability still scales as (τ/N)^{N_m}; or introduce detectives and test whether winning curves for different compositions still refuse to collapse under the same one-parameter Nc rescaling.

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

If this is right

  • Without detectives, larger populations favor the good faction once N exceeds an empirical crossover Nc that depends on initial mafia size.
  • Early mafia extinction risk is controlled mainly by initial mafia count and total population via a simple power of τ/N when the mafia is rare.
  • A few perfectly informed detectives sharply cut mafia survival times and change the effective exponent of the cumulative extinction curves.
  • With detectives present, good-faction composition becomes an extra control variable, so one-parameter population rescaling no longer collapses win probabilities.
  • Macroscopic absorption statistics of hidden-role games can be linked to microscopic accusation histories without modeling full dialogue.

Where Pith is reading between the lines

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

  • The same suspicion-accumulation skeleton could be reused to test whether other social-deduction or rumor settings show an analogous dilute-regime power and finite-size collapse.
  • If real human accusation patterns differ systematically from the stipulated pair rules, fitting those kernels from play logs would be the natural next experiment before claiming universality.
  • The detective-as-relevant-field picture suggests that even small fractions of trusted verifiers may move hidden-role contests out of the pure no-detective scaling class.
Share X Bluesky LinkedIn Reddit HN

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 paper introduces an agent-based Mafia/Werewolf model in which the day-phase elimination is mediated by cumulative, player-specific suspicion scores updated through role-dependent pairwise interactions, rather than by uniform random voting. The dynamics has two absorbing states (mafia extinction or mafia–good parity). The authors derive a mean-field benchmark: under a random-execution approximation with independent mafia death times in the dilute regime (N_m ≪ N), the cumulative probability of mafia extinction behaves as F(τ) ~ (τ/N)^{N_m} at early times. Monte Carlo simulations (N_run = 2×10^5) are reported to confirm this for a dilute case (N=512, N_m=2, fitted α≃2.03 vs predicted 2). The good-faction winning probability as a function of N is sigmoidal, and curves for different N_m approximately collapse when N is rescaled by an empirical crossover scale N_c defined by Π_g = 0.5; with perfectly informed detectives the same one-parameter collapse fails. The authors frame the crossover as a nonequilibrium phase transition.

Significance. If the claims hold up, the paper contributes a clean, minimal agent-based link between microscopic accusation histories and macroscopic absorption statistics in hidden-role games, a niche currently split between exact-but-rigid mathematical treatments and heavy AI-gameplay studies. Specific strengths: the mean-field prediction α = N_m is a genuinely falsifiable, parameter-free relation; the Monte Carlo protocol is transparent and well-powered; the absorbing-state framing is sound; and the negative result (detectives break the one-parameter collapse) is informative about when simple finite-size rescaling should fail. The work is incremental over Migdał (2010) on the analytic side, but the suspicion-mediated day phase is new and the model is extensible.

major comments (3)
  1. [§IV, Fig. 3 and Eq. (38)] The central quantitative claim — that the cumulative extinction probability approaches (τ/N)^{N_m} 'in agreement with simulations as N_m/N decreases' — is supported by essentially two fits: α≃4.7 vs predicted 4 at N_m=4, N=256 (a regime the authors themselves note is not dilute), and α≃2.03 at N_m=2, N=512, τ_max=50. The fitted exponent drifts monotonically with the fit window (2.03 → 2.10 as τ_max goes 50 → 100), so the agreement is window-sensitive. To establish the claim the paper needs a systematic scan: α(N_m, N) for several (N_m, N) pairs at fixed small N_m/N, plus a window-stability analysis (e.g., local-slope plots or extrapolation of α as τ_max/N → 0). As it stands, the evidence distinguishes 'consistent with' from 'demonstrates' only at a single point.
  2. [§IV, Fig. 4 inset and surrounding text] The collapse is assessed only visually, and the collapse coordinate N_c is itself spline-estimated from each curve at exactly the feature being aligned (Π_g = 0.5). Any family of smooth, ordered sigmoids will align at their median crossings under this normalization, so some degree of collapse is expected by construction. The claim would be substantially strengthened by (i) a quantitative collapse measure (pooled residuals from an interpolated master curve, compared against a null such as individually shifted sigmoids), (ii) bootstrap uncertainties on N_c, and (iii) a report of the N_c(N_m) dependence — which would also allow a direct comparison with the Braverman–Etesami–Mossel result that the balanced regime sits at N_m ~ sqrt(N), i.e., N_c ~ N_m^2. This last test is available within the authors' existing simulation pipeline and would convert an empirical observation into a checked pred
  3. [§V] The Discussion elevates the N_c crossover to a 'non-equilibrium phase transition', a 'unique universality class', and the detective density to a 'relevant field perturbation'. The data shown are smooth sigmoidal crossovers with no singular behavior, no critical exponents, and no scaling function beyond the single-parameter median normalization. This terminology is not supported by the evidence and misplaces the result relative to the absorbing-state literature cited (Refs. 14–17). The language should be tempered to 'finite-size crossover' (which the abstract itself uses), or else substantiated with a genuine scaling analysis (e.g., sharpening of the transition width with N, data collapse with a second scaling variable).
minor comments (7)
  1. [Fig. 1 caption] Fig. 1 caption ends with 'N_d^{(0)}.' — the value appears to be missing (presumably 0).
  2. [Throughout, esp. §IV] Terminology shifts across the text and captions: players are variously 'citizens', 'peasants', 'villagers', 'mobsters', 'werewolves', 'predators', and 'spreaders of fake news'. Fig. 2's caption and text use n_w and n_d where the rest of the paper uses N_m and N_d. A single consistent notation and vocabulary would help.
  3. [§IV, Eq. (39)] Eq. (39) reports a rank-plot exponent t_survival ~ rank^{0.2} for N_m=4. The relation between this rank exponent and the cumulative exponent α is not explained; since rank ordering is the inverse CDF, one expects an exponent ~1/α, which for α≃4.7 gives ~0.21 — worth stating explicitly, as it would make Fig. 2(a) an independent consistency check rather than a loose remark.
  4. [§V] Typo: 'sociocompational' in the first paragraph of §V (presumably 'sociocomputational'). Also 'Mafia Party game' is inconsistent with usage elsewhere.
  5. [§II.C] The simulation code does not appear to be made available. Given that the results are purely numerical with a clean protocol (N_run = 2×10^5), releasing the code would materially improve reproducibility and is standard practice in this field.
  6. [§II.B, Eqs. (13)–(15)] The special-case rule for A(t)=0 after Eq. (15) and the half-weight unpaired-agent rule in Eq. (13) are sensible but ad hoc; a brief remark on their quantitative impact (how often A=0 occurs in practice; odd-population frequency) would help readers assess robustness.
  7. [§III] The overlap with Migdał (Ref. 19) is disclosed honestly at the start of §III, which is commendable; consider also briefly stating what, if anything, differs technically (e.g., treatment of the parity absorbing boundary), since 'alternative formulation' is otherwise vague.

Circularity Check

1 steps flagged

No load-bearing circularity: early-time power law is an independent random-execution benchmark compared to simulation; N_c collapse is standard operational finite-size scaling, not a forced prediction.

specific steps
  1. fitted input called prediction [Sec. IV, Fig. 4a inset; text on N_c extraction]
    "the inset of Fig. 4a displays the same probability plotted against the scaled system size N(0)/N_c, where N_c represents the critical characteristic population initial size for which Π̂_p = 0.5. Remarkably, this rescaling leads to an excellent data collapse of all curves onto a master curve. ... To accurately extract the values of N_c for each independent curve where Π̂_p = 0.5, a cubic spline interpolation scheme was implemented over the discrete simulation data points."

    N_c is estimated from each winning-probability curve itself (the Π_g=0.5 crossing) and then used as the sole horizontal scale for that same curve. Agreement at the half-point is therefore automatic; only residual shape mismatch could prevent collapse. This is ordinary finite-size scaling practice and does not force the full master-curve claim, but it is a mild fitted-coordinate step rather than an a-priori predicted scale.

full rationale

The paper’s central analytic claim is the dilute/early-time cumulative extinction law F(τ)∼(τ/N)^{N_m}. Section III derives this from a deliberately simplified random-execution mean-field (uniform day kills, independent mafia death times when N_m≪N), obtaining q(t)=√(1−2t/N) and then Pr(t_survival≤τ)=[1−q(τ)]^{N_m}≈(τ/N)^{N_m}. That derivation does not use the agent-based suspicion scores, nor is it fitted to the Monte Carlo output it is later compared against (Figs. 2–3). Overlap with Migdał is acknowledged; the benchmark is externally falsifiable and is only partially confirmed (clean α≈2.03 in one dilute window). The winning-probability collapse defines N_c operationally as the spline-estimated N_0 where Π̂_g=0.5 and rescales N_0/N_c—standard one-parameter finite-size scaling that forces agreement only at the half-crossing, not full curve collapse. Detectives breaking the same rescaling is an independent negative control. Self-citations (e.g. the boxers-game paper) are motivational, not uniqueness or ansatz load-bearing. No step reduces a claimed first-principles prediction to its own fitted input by construction. Score 1 only for the mild, conventional use of data-defined N_c as the collapse coordinate.

Axiom & Free-Parameter Ledger

2 free parameters · 8 axioms · 2 invented entities

The paper’s load-bearing content is a stipulated microscopic interaction table plus standard absorbing-state Monte Carlo. Analytic content in the dilute limit re-derives a known pure-death approximation. The only fitted objects used in the headline collapse are the per-curve N_c values. No new physical constants; the suspicion scores and role-pair kernels are invented model entities without external calibration.

free parameters (2)
  • N_c (crossover population where Π_g=0.5) = curve-dependent; read off where estimated good-win probability equals 1/2
    Extracted per initial-mafia curve by cubic-spline interpolation on simulation win probabilities; used as the sole collapse coordinate in Fig. 4a. No closed-form prediction for N_c(N_m).
  • Early-time fit window τ_max = τ_max=50 primary; sensitivity reported
    Exponent checks depend on the chosen upper fit bound (τ_max=50,70,90,100), with α drifting from 2.03 to 2.10; window is chosen by hand.
axioms (8)
  • ad hoc to paper Day elimination probability equals normalized cumulative suspicion p_i=a_i/A (or uniform if A=0).
    §II.B Eq. (15); defines the macroscopic link from microscopic accusations to absorption. Not derived from rational voting or empirical play.
  • ad hoc to paper Role-dependent pairwise suspicion kernels (mafia smear good; detectives always correctly accuse mafia; civilians random; mafia–detective split 1/2).
    §II.B bullet list Eqs. (8)–(13). Entire microscopic engine; alternative kernels would change statistics.
  • domain assumption Suspicion scores are permanent until the agent dies; no decay, private information, or strategic withholding.
    §II opening and Eq. (3)–(4). Makes the process history-dependent and Markovian only in the full suspicion configuration.
  • domain assumption In the dilute limit N_m≪N, mafia death times may be treated as independent, yielding binomial extinction statistics.
    §III Eqs. (26)–(34); standard mean-field independence closure, acknowledged as approximate.
  • domain assumption Night kill is uniform over the surviving good faction; mafia never dies at night.
    §II.A Eq. (7); standard Mafia-rule idealization used by Braverman/Migdał baselines.
  • domain assumption Absorbing outcomes: good wins at N_m=0; mafia wins at N_m≥N_g, checked after each elimination.
    §II Eqs. (5)–(6); canonical termination rules of the party game.
  • domain assumption Well-mixed random pairing each day; no spatial or network structure.
    §II.B; enables mean-field comparison and simplifies Monte Carlo.
  • ad hoc to paper Detectives are perfectly informed and always direct informative accusations at true surviving mafia; no private investigation memory is stored.
    §II.B coarse-grained detective paragraph. Strong idealization that drives the shortened survival times and broken collapse.
invented entities (2)
  • Player-specific cumulative public suspicion scores a_i(t) no independent evidence
    purpose: Mediate history-dependent day eliminations so that identical faction counts can have different transition probabilities.
    Core model invention in §II; replaces uniform random voting used in prior mathematical Mafia models.
  • Empirical finite-size crossover scale N_c no independent evidence
    purpose: Organize winning-probability curves for different N_m,0 onto an approximate master curve when detectives are absent.
    Defined operationally in §IV from Π_g=0.5; not derived from a renormalization-group or closed-form argument.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Cumulative suspicion and absorption dynamics in an agent-based Mafia game." pith.science (2026). https://pith.science/paper/CMOUSGJB

@misc{pith2026260724980,
  author       = {Pith},
  title        = {Pith review of: Cumulative suspicion and absorption dynamics in an agent-based Mafia game},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMOUSGJB}},
  note         = {Machine review of arXiv:2607.24980}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The Mafia game, also known as Werewolf, describes a competition between a coordinated minority whose identities are concealed and a larger good faction composed primarily of uninformed civilians attempting to identify and eliminate them. We introduce an agent-based formulation in which the daytime decision is not represented by uniform random voting. Instead, randomly paired agents modify player-specific public suspicion scores according to their roles, and one surviving player is subsequently eliminated with probability proportional to their accumulated suspicion. These scores persist throughout the game, producing a history-dependent stochastic process with two competing absorbing outcomes: elimination of the mafia or numerical parity between the mafia and the good faction. We investigate the effects of population size, initial mafia size, and the presence of perfectly informed detectives through extensive Monte Carlo simulations. In the absence of detectives, a random-execution approximation predicts that, in the dilute-mafia and early-time regime, the cumulative probability of mafia extinction behaves as $F(\tau)\sim (\tau/N)^{N_m}$, in agreement with simulations as $N_m/N$ decreases. The winning-probability curves also exhibit an empirical finite-size crossover characterized by a population scale $N_c$. Rescaling the population by $N_c$ approximately collapses the curves obtained for different initial mafia populations. Perfectly informed detectives substantially shorten mafia-survival times and introduce an additional dependence on population composition for which the same one-parameter rescaling is insufficient. The model provides a minimal connection between microscopic histories of accusation and the macroscopic absorption statistics of hidden-role games.

Figures

Figures reproduced from arXiv: 2607.24980 by Eduardo Velasco Stock, Roberto da Silva.

Figure 1
Figure 1. Figure 1: FIG. 1: The typical evolution of several seeds showing the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Check of the mean-field behavior of the cumula [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Survival times for the mafia. (a) Rank plot of the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Estimated conditional distribution of villagers win [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Review history (2 revisions) →

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

30 extracted references · 3 linked inside Pith

  1. [1]

    Statistical physics of social dynamics.Reviews of Modern Physics, 81(2):591–646, 2009

    Claudio Castellano, Santo Fortunato, and Vittorio Loreto. Statistical physics of social dynamics.Reviews of Modern Physics, 81(2):591–646, 2009

    2009

  2. [2]

    Springer, New York, 2012

    Serge Galam.Sociophysics: A Physicist’s Modeling of Psycho-political Phenomena. Springer, New York, 2012

    2012

  3. [3]

    A model for spatial conflict.Biometrika, 60(3):581–588, 1973

    Peter Clifford and Aidan Sudbury. A model for spatial conflict.Biometrika, 60(3):581–588, 1973

    1973

  4. [4]

    Holley and Thomas M

    Richard A. Holley and Thomas M. Liggett. Ergodic theo- rems for weakly interacting infinite systems and the voter model.The Annals of Probability, 3(4):643–663, 1975

    1975

  5. [5]

    The dissemination of culture: A model with local convergence and global polarization.Journal of Conflict Resolution, 41(2):203–226, 1997

    Robert Axelrod. The dissemination of culture: A model with local convergence and global polarization.Journal of Conflict Resolution, 41(2):203–226, 1997

    1997

  6. [6]

    Stochasticrumours.Jour- nal of the Institute of Mathematics and its Applications, 1(1):42–55, 1965

    D.J.DaleyandD.G.Kendall. Stochasticrumours.Jour- nal of the Institute of Mathematics and its Applications, 1(1):42–55, 1965

    1965

  7. [7]

    Maki and Maynard Thompson.Mathematical Models and Applications: With Emphasis on the Social, Life, and Management Sciences

    Daniel P. Maki and Maynard Thompson.Mathematical Models and Applications: With Emphasis on the Social, Life, and Management Sciences. Prentice-Hall, Engle- wood Cliffs, NJ, 1973

    1973

  8. [8]

    Minority opinion spreading in random ge- ometry.The European Physical Journal B, 25(4):403– 406, 2002

    Serge Galam. Minority opinion spreading in random ge- ometry.The European Physical Journal B, 25(4):403– 406, 2002

    2002

  9. [9]

    P. L. Krapivsky and S. Redner. Dynamics of majority rule in two-state interacting spin systems.Physical Re- view Letters, 90(23):238701, 2003

    2003

  10. [10]

    Does a single zealot affect an infinite group of voters?Physical Review Letters, 91(2):028701, 2003

    Mauro Mobilia. Does a single zealot affect an infinite group of voters?Physical Review Letters, 91(2):028701, 2003

    2003

  11. [11]

    David M. J. Lazer, Matthew A. Baum, Yochai Benkler, Adam J. Berinsky, Kelly M. Greenhill, Filippo Menczer, Miriam J. Metzger, Brendan Nyhan, Gordon Pennycook, David Rothschild, Michael Schudson, Steven A. Sloman, Cass R. Sunstein, Emily A. Thorson, Duncan J. Watts, and Jonathan L. Zittrain. The science of fake news.Sci- ence, 359(6380):1094–1096, 2018

    2018

  12. [12]

    The spread of true and false news online.Science, 359(6380):1146– 1151, 2018

    Soroush Vosoughi, Deb Roy, and Sinan Aral. The spread of true and false news online.Science, 359(6380):1146– 1151, 2018

    2018

  13. [13]

    boxers game

    Roberto da Silva and Henrique A. Fernandes. When and how much the altruism impacts your privileged informa- tion? proposing a new paradigm in game theory: The “boxers game”.Physica A: Statistical Mechanics and its Applications, 503:283–292, 2018

    2018

  14. [14]

    Non-equilibrium critical phenomena and phase transitions into absorbing states.Advances in Physics, 49(7):815–958, 2000

    Haye Hinrichsen. Non-equilibrium critical phenomena and phase transitions into absorbing states.Advances in Physics, 49(7):815–958, 2000

    2000

  15. [15]

    Universality classes in nonequilibrium lat- tice systems.Reviews of Modern Physics, 76(3):663–724, 2004

    Géza Ódor. Universality classes in nonequilibrium lat- tice systems.Reviews of Modern Physics, 76(3):663–724, 2004

    2004

  16. [16]

    Theoretical and Mathematical Physics

    Malte Henkel, Haye Hinrichsen, and Sven Lübeck.Non- Equilibrium Phase Transitions: Volume 1: Absorb- ing Phase Transitions. Theoretical and Mathematical Physics. Springer, Dordrecht, 2008

    2008

  17. [17]

    Theoretical and Mathe- matical Physics

    Malte Henkel and Michel Pleimling.Non-Equilibrium Phase Transitions: Volume 2: Ageing and Dynamical Scaling Far from Equilibrium. Theoretical and Mathe- matical Physics. Springer, Dordrecht, 2010

    2010

  18. [18]

    Mafia: A theoretical study of players and coalitions in a partial information environment.The Annals of Applied Probability, 18(3):825–846, 2008

    Mark Braverman, Omid Etesami, and Elchanan Mossel. Mafia: A theoretical study of players and coalitions in a partial information environment.The Annals of Applied Probability, 18(3):825–846, 2008

    2008

  19. [19]

    A mathematical model of the mafia game

    Piotr Migdał. A mathematical model of the mafia game. arXiv preprint arXiv:1009.1031, 2010

    Pith/arXiv arXiv 2010

  20. [20]

    S. T. Wang. Optimal strategy in werewolf game: A game theoretic perspective.arXiv preprint arXiv:2408.17177, 2024

    arXiv 2024

  21. [21]

    Keeping the story straight: A comparison of commitment strategies for a social deduction game

    Markus Eger and Chris Martens. Keeping the story straight: A comparison of commitment strategies for a social deduction game. InProceedings of the Fourteenth AAAI Conference on Artificial Intelligence and Interac- tive Digital Entertainment, 2018

    2018

  22. [22]

    Exploring large language models for communication games: An empiri- cal study on werewolf.arXiv preprint arXiv:2309.04658, 2023

    Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiao- long Wang, Weidong Liu, and Yang Liu. Exploring large language models for communication games: An empiri- cal study on werewolf.arXiv preprint arXiv:2309.04658, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Werewolf arena: A case study in llm evaluation via social deduc- tion.arXiv preprint arXiv:2407.13943, 2024

    Suma Bailis, Jane Friedhoff, and Feiyang Chen. Werewolf arena: A case study in llm evaluation via social deduc- tion.arXiv preprint arXiv:2407.13943, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Findingdeceivers in social context with large language models and how to find them: the case of the mafia game.Scientific Reports, 14(1):30946, 12 2024

    ByunghwaYooandKyung-JoongKim. Findingdeceivers in social context with large language models and how to find them: the case of the mafia game.Scientific Reports, 14(1):30946, 12 2024

    2024

  25. [25]

    Putting the con in context: Identifying deceptive actors in the game of mafia, 2022

    Samee Ibraheem, Gaoyue Zhou, and John DeNero. Putting the con in context: Identifying deceptive actors in the game of mafia, 2022

    2022

  26. [26]

    Detection of deception in the mafia party game

    Sergey Demyanov, James Bailey, Kotagiri Ramamoha- narao, and Christopher Leckie. Detection of deception in the mafia party game. InProceedings of the 2015 ACM on International Conference on Multimodal Interaction, ICMI ’15, page 335–342, New York, NY, USA, 2015. As- sociation for Computing Machinery

    2015

  27. [27]

    Rehg, and Diyi Yang

    Bolin Lai, Hongxin Zhang, Miao Liu, Aryan Pari- ani, Fiona Ryan, Wenqi Jia, Shirley Anugrah Hayati, James M. Rehg, and Diyi Yang. Werewolf among us: A multimodal dataset for modeling persuasion behaviors in social deduction games, 2022. 9

    2022

  28. [28]

    Cambridge University Press, Cambridge, 2001

    Sidney Redner.A Guide to First-Passage Processes. Cambridge University Press, Cambridge, 2001

    2001

  29. [29]

    Fisher and Michael N

    Michael E. Fisher and Michael N. Barber. Scaling the- ory for finite-size effects in the critical region.Physical Review Letters, 28(23):1516–1519, 1972

    1972

  30. [30]

    International Series of Monographs on Physics

    H.EugeneStanley.Introduction to Phase Transitions and Critical Phenomena. International Series of Monographs on Physics. Clarendon Press, Oxford, 1971

    1971

This paper was first reviewed by grok-4.5 on July 31, 2026.