{"id":"8592ca0a-047e-414c-a453-39a0ec65112d","arxiv_id":"2504.18039","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MultiMind combines facial and vocal emotion cues with a learned theory-of-mind belief model and Monte Carlo tree search to make Werewolf agents more persuasive and less suspected.","lead":"This paper builds an AI agent for the social deduction game One Night Ultimate Werewolf that reads facial expressions and vocal tones in addition to text, then plans its statements to reduce suspicion against itself. A generalist reader might care because it tests whether language agents can combine perception, belief modeling, and strategic communication in human-like social settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ToM model's ground-truth beliefs are players' public statements, not private beliefs; MCTS optimizes against predicted stated suspicion, and the paper never shows this proxy tracks actual votes, so the reported win-rate gains may not be caused by accurate belief modeling.","rationale":"We read the paper as a systems contribution: a ToM model trained from self-play and human data, an MCTS planner that uses it as a reward model, and a multimodal perceiver/Actor pipeline, evaluated in mixed-agent and human games. The headline win-rate numbers (49.8% vs 42.3% best baseline in Table 1) and the ablations (Tables 2-3) are real evidence that the full system plays well; we are not disputing that. The load-bearing issue is whether the ToM model, the component that makes the planning 'theory-of-mind', actually models what opponents believe. The training labels in Eq. (15) are self-reported suspicions obtained by prompting agents after each speech. In a social deduction game, such reports are strategic public statements and can be false. The MCTS objective (Eq. 11) therefore minimizes predicted stated suspicion, not private belief or votes. The paper never validates the mapping from B_t to actual voting behavior, either on training opponents or held-out ones. This is a more fundamental concern than the reader's generalization worry: even a perfectly generalized ToM model would be optimizing the wrong target if the labels are not beliefs. The reader's proposed 'direct evaluation of ToM belief accuracy on held-out opponents' would partially address this, but the test must use actual votes as the ground truth, not another self-report. We therefore recommend keeping the CONDITIONAL verdict: the system works, but the causal claim about ToM-driven planning is not yet established. We also note the absence of significance testing and the human-study asymmetry, but those are secondary to the reward mis-specification.","tokens_in":17450,"tokens_out":9444,"duration_ms":93361,"concrete_test":"Run 100 held-out mixed-agent games with the same opponents (Gemini-2.0-Flash and baselines), logging at each decision point the ToM model's predicted B_t[j,i] for every player j and target i, plus the actual vote cast by j at the end of the day phase. Compute the Spearman rank correlation between predicted suspicion and vote target. If the correlation is weak (<0.3), the reward in Eq. (11) is not aligned with the actual objective, and the win-rate gains require an alternative explanation. Additionally, retrain the ToM model using each player's final vote (one-hot target) instead of the stated-suspicion labels in Eq. (15), keeping all else fixed; if the vote-trained variant achieves comparable or better win rate and lower average votes in the same mixed-agent evaluation, the stated-suspicion proxy is not responsible for the planner's gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the ToM model's predicted belief matrix B_t represents what opponents actually believe, so that the MCTS planner in Eqs. (8)-(11) minimizes real suspicion. This assumption fails at the training-signal level: the ground-truth belief matrix in Eq. (15) is constructed by prompting LLM agents to state their 'werewolf suspicions' after each utterance. These are public, strategic statements in a social deduction game, not private beliefs; players can and do lie. The planner's reward in Eq. (11) is the negative sum of the ToM model's output B_{t+1}[j,i], so it minimizes predicted stated suspicion, not predicted votes or true belief. The paper provides no evidence that stated suspicion predicts voting: there is no calibration or correlation check between B_t and the votes actually cast by held-out opponents (Gemini-2.0-Flash and the baseline agents). If the proxy is mis-specified, the planner optimizes the wrong objective, and the observed win-rate advantage (Table 1) could come from other components, such as the LLM Actor's longer, more coherent statements, the extra planning compute, or the text emotion labels, rather than from accurate ToM. Table 2 shows the ToM model outperforming an LLM reasoner at matched compute, but both are evaluated against the same proxy-based objective, so this comparison does not resolve the mis-specification. A direct alignment test is needed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MultiMind augments an LLM agent for One Night Ultimate Werewolf (ONUW) with multimodal emotion perception (facial-expression and vocal-tone labels converted to text), a trained Transformer-based Theory of Mind (ToM) model that predicts a belief matrix B_t of each player's suspicion toward each other player, and an MCTS planner that selects the next action/emotion sequence to minimize predicted suspicion against the agent. The ToM model is trained on LLM self-play data and then fine-tuned on human gameplay data. Experiments report higher win rates and lower vote counts for MultiMind than four baselines in mixed-agent games (Table 1), ablations comparing MCTS with random/DFS/BFS planning (Table 3) and the compact ToM reasoner with an LLM reasoner (Table 2), plus an 8-participant human study (Table 4). The central claim is that adding multimodal input, second-order ToM modeling, and MCTS planning makes the agent win more often and receive fewer votes.","tokens_in":17632,"tokens_out":8155,"duration_ms":83813,"significance":"If the results hold, the paper provides a practical demonstration of integrating multimodal perception, second-order ToM reasoning, and planning in an LLM social-deduction agent, with an open-source implementation and a sensible two-stage training scheme. The agent-agent results show a consistent direction across multiple tables, and the comparison of the lightweight ToM reasoner against an LLM reasoner under a matched planning budget is useful evidence for the efficiency of the proposed architecture. The main source of uncertainty is whether the mechanism attributed to ToM is actually what is being measured: the MCTS objective is defined in terms of the ToM model's predictions, but those predictions are trained on LLM self-reported suspicions that are not validated against private beliefs or actual votes. The multimodal claims are also only partially supported, because agent-agent games use self-generated text emotion labels rather than real perception, and the human study confounds emotion-label communication with the overall agent design.","major_comments":[{"comment":"The ground-truth labels for the ToM model are the LLM agents' self-reported suspicions after each utterance. These self-reports are not shown to correspond to the agents' private beliefs or to their eventual votes, and in a social-deduction setting an agent may answer strategically. Because the MCTS objective in Eq. (11) minimizes the ToM model's predicted suspicion B_{t+1}[j,i], the planner could be optimizing against a model of stated suspicion that is mis-specified with respect to the actual voting behavior of held-out opponents. The paper reports no calibration or correlation check between B_t and the votes cast by Gemini-2.0-Flash or the baseline agents, and the human-data fine-tuning additionally imputes missing intention annotations with GPT-4o (Section 3.6.2), another LLM-generated label source. Please add a direct alignment test, for example predicting vote distributions from B_t and comparing them with observed votes on held-out games, and report the correlation for opponents not used in training.","section":"Section 3.6.1, Eq. (15) and Section 3.4, Eq. (11)"},{"comment":"In the agent-agent experiments, e_face and e_tone are text labels written by the LLM actors themselves rather than outputs of the audio/video perception pipeline, so Tables 1-3 do not test multimodal perception, only the effect of adding an explicit emotion-label communication channel. In the human study, only MultiMind's Actor emits such labels, so the lower vote count for MultiMind in Table 4 is confounded with the presence of this extra channel. Please either restrict the claims to 'text emotion labels' for agent experiments, add a control baseline that emits random or fixed emotion labels, or run an ablation in which MultiMind does not append emotion labels.","section":"Section 4.1.2 and Section 3.5"},{"comment":"The paper reports no confidence intervals, standard errors, or significance tests for any win-rate or vote-count comparison. For example, the 61.0% versus 66.0% overall win rates in Table 3 are based on 50 games per condition, which gives a standard error of roughly 7 percentage points for each rate; differences of this size are compatible with sampling noise. Please report bootstrap confidence intervals and appropriate tests for all headline comparisons, including the mixed-agent results in Table 1 and the human study in Table 4.","section":"Section 4.2, Table 1 and Section 4.3, Tables 2-3"}],"minor_comments":[{"comment":"The sentence claiming to be 'the first to develop a framework that integrates multimodal information into SDG agents' should be qualified, since Section 2.2 only surveys retrospective multimodal analyses; if no active multimodal SDG agents exist, a sentence stating that explicitly would help.","section":"Section 1"},{"comment":"The maximum action-sequence depth |A_{t+1}|_max = 3 is asserted without justification; please report sensitivity to this hyperparameter or at least discuss why three statements is sufficient.","section":"Section 3.4"},{"comment":"Panel (b) appears to compare win rates against each baseline, but the caption does not define the grouping or the number of games per bar; please add a descriptive caption and error bars.","section":"Section 4.3, Figure 5"},{"comment":"The use of GPT-4o to fill missing intention annotations should be validated on a small subset or acknowledged as a limitation, since it introduces an additional LLM-labeled target into the human fine-tuning set.","section":"Section 3.6.2"},{"comment":"The table formatting is corrupted in the submitted PDF (e.g., '2900.8337.2' and '861.05 70.93380.83'); please ensure all numeric columns are clearly separated.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for the multimedia venue because of the multimodal emotion component, but that component is the least validated part of the system: the agent-agent experiments only use self-generated text emotion labels, and the human study confounds the emotion-label channel with the rest of the design. I would encourage the editor to require the proposed calibration and control experiments before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhang et al. build MultiMind, an ONUW agent that fuses facial/vocal emotion labels into a ToM belief model and uses MCTS to plan suspicion-minimizing statements. The package is real: code released, 5,300 self-play games, ablations against alternative reasoners and planners, and a human study. The central result—a coherent agent that wins more and gets voted less in mixed-agent games—is plausibly solid. The ToM+MCTS combination is a new integration, and the ablations show the ToM reasoner beats an LLM reasoner at equal compute, and MCTS beats DFS/BFS/random with equal calls.\n\nMy main concern matches the stress-test: the ToM model's ground-truth beliefs are LLM self-reported suspicions, which are public strategic statements, not private beliefs. The MCTS reward minimizes predicted suspicion, and there is no check that B_t predicts actual votes from held-out opponents. Table 2 compares ToM vs LLM reasoner on the same proxy objective, so it doesn't resolve this. The win-rate gains could come from the actor's longer or more coherent text or extra planning compute. This is a proxy risk, not a fatal flaw—the system works as a suspicion-minimizer against stated suspicion—but the paper overclaims 'belief' modeling.\n\nSecond, no error bars or significance tests anywhere; 400 mixed-agent games and 30-80 per cell in Figure 4. LLM agents are stochastic, so this is a real evidentiary gap. Bootstrap confidence intervals or pairwise tests would fix it. Third, the multimodal claim is weaker than the abstract implies. In agent-agent games, emotion labels are text generated by the agents themselves, so the multimodal pipeline is not actually exercised. The human study does use real audio/video, but it's eight people and MultiMind can append emotion text labels to its messages while baselines cannot—a confound.\n\nThe paper is a solid systems contribution and the authors are honest about limitations in Section 4.5. The remaining gap is evidentiary: they need a direct test of ToM belief accuracy (calibration against votes) and a matched ablation that removes the text emotion-label channel in the human study. With that, the central claim would be much better supported. As is, I'd send it to review—it deserves referee time—but my own verdict would be 'revise with statistics and a belief-alignment check.' Who's it for: LLM-agent and social-deduction researchers, plus anyone building ToM modules for games. I'd cite it if I worked on SDG agents, but I'd hedge about the belief proxy.","headline":"A well-engineered ONUW agent with a genuinely new ToM+MCTS integration whose win-rate gains are real but whose belief model rests on a proxy that needs a direct calibration check.","tokens_in":18289,"tokens_out":1747,"would_cite":true,"duration_ms":16615,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Multimodal perception and a learned theory-of-mind model let a One Night Ultimate Werewolf agent win 49.8 percent of mixed-agent games and attract the fewest human votes.","keywords":["social deduction games","One Night Ultimate Werewolf","theory of mind","multimodal emotion recognition","Monte Carlo Tree Search","large language model agents","belief modeling","suspicion minimization"],"falsifier":"Set MultiMind against a held-out set of opponent agents, record each opponent's true suspicion after every statement, and compare it with the theory-of-mind model's predicted belief matrices; if the predictions are no better than chance, or if the search-chosen utterances do not reduce the opponents' measured suspicion below randomly chosen utterances, the paper's central explanation collapses.","tokens_in":17080,"feed_emoji":"🐺","tokens_out":11585,"duration_ms":100574,"temperature":0.7,"pith_summary":"This paper argues that a social-deduction agent plays better when it can see and hear its opponents, not just read their words. Its MultiMind system attaches facial-expression and vocal-tone labels to each statement, learns a theory-of-mind model that outputs how much each player suspects every other player of being the werewolf, and uses Monte Carlo Tree Search to choose statements that lower suspicion directed at itself. In 400 mixed-agent games the agent wins 49.8 percent of the time, versus 42.3 percent for the best text-only baseline, and in a study with human players it receives the fewest votes. The paper positions this as a step toward LLM agents that engage in human-like social reasoning across multimodal domains.","feed_headline":"Werewolf agent wins 49.8% of games by reading faces and tones","feed_subtitle":"It models who suspects whom and plans statements that draw the fewest votes against human players.","key_machinery":"The load-bearing object is the learned theory-of-mind model in the Reasoner: a 25.4-million-parameter causal Transformer that turns the game history—action triplets plus facial and vocal emotion labels—into a belief matrix $B_t[i,j]$ at each time step. This matrix converts 'what does the other player think?' from a prompt-level intuition into a callable reward signal. The Planner wraps it in Monte Carlo Tree Search: for each candidate statement the model predicts a new belief matrix, and the reward $-\\sum_{j \\neq i} B_{t+1}[j,i]$ propagates back through the tree, so the agent selects utterances that reduce the predicted suspicion pointed at it.","core_discovery":"MultiMind's central claim is that in One Night Ultimate Werewolf, suspicion can be treated as a measurable, learnable quantity. Every player statement is parsed into action triplets (speaker, action, target) and each utterance is tagged with one of eight facial-emotion and one of eight vocal-tone labels; a small causal Transformer consumes this sequence and outputs a belief matrix $B_t[i,j]$, the probability that player $i$ believes player $j$ is the werewolf. The planning loop then searches over candidate next utterances with Monte Carlo Tree Search, scoring each by the negative sum of predicted suspicion directed at the agent after that utterance. Trained first on LLM self-play and fine-tuned on human game recordings, the pipeline wins 70.9 percent of games as the werewolf and 44.4 percent as a villager across 400 mixed-agent games (49.8 percent overall), and against human players it draws the fewest votes (0.87 average per game, 0.19 from humans). The paper takes this as evidence that multimodal emotion and second-order belief modeling improve active gameplay, not just retrospective analysis.","pith_inferences":["A natural next test is whether the same perceiver-plus-belief-matrix-plus-search recipe transfers to other social games or negotiations, where the target is not a single werewolf probability but a richer attitude structure; the paper's equations only assume a suspicion distribution over players.","Because emotions are collapsed to eight coarse categories, the reported gains may understate what continuous facial and prosodic features could contribute to belief prediction, or the emotion classifiers' errors may be the current bottleneck.","The suspicion-minimization objective could reward bland or self-effacing statements that reduce votes at the cost of information useful to teammates; a testable extension is to add an information-gain or team-utility term and measure whether village win rate rises further.","The theory-of-mind model is used as a fixed oracle inside the search; an iterative planner that updates its beliefs about opponents as the game unfolds would test whether the fixed-oracle assumption is the main limit."],"forward_implications":["Multimodal emotion labels earn their place in the model: removing either facial or vocal emotion raises the theory-of-mind validation loss, and removing both lowers win rates against every baseline.","The learned belief model is a viable substitute for asking an LLM to infer beliefs: at equal MCTS iterations it matches the LLM-based Reasoner while being far faster, freeing compute for more search.","The search structure itself contributes to the gain: MCTS beats random sampling, depth-first search, and breadth-first search over the same number of belief-model evaluations.","The performance carries over to an LLM backend that was not used in training-data generation, suggesting the suspicion model generalizes beyond the models that produced its training games."],"supporting_citations":[{"why":"Provides the 5-player One Night Ultimate Werewolf environment and the LLM-instructed and RL-instructed baselines that MultiMind is compared against.","marker":"[12]"},{"why":"Supplies the human gameplay videos, transcripts, timestamps, and intention annotations used to fine-tune the theory-of-mind model and to study multimodal cues.","marker":"[13]"},{"why":"Supplies the ReAct baseline that prompts the LLM directly with raw observations.","marker":"[39]"},{"why":"Supplies the facial-expression classifier that turns video frames into eight emotion labels for human opponents.","marker":"[5]"},{"why":"Supplies the audio model used to transcribe speech and classify vocal tone into emotion labels.","marker":"[7]"},{"why":"Supplies the causal Transformer architecture on which the theory-of-mind model is built.","marker":"[30]"},{"why":"Supplies an LLM used to annotate unlabeled human segments and as one of the backends in self-play training.","marker":"[22]"},{"why":"One of the LLM backends that generated self-play games for training the theory-of-mind model.","marker":"[9]"},{"why":"One of the LLM backends that generated self-play games for training the theory-of-mind model.","marker":"[25]"}],"fun_headline_variants":["AI Werewolf agent reads faces and tones to dodge suspicion","Multimodal AI Werewolf agent predicts suspicion, wins 49.8% of games","Faces, tones, and theory of mind: AI Werewolf agent plans around suspicion","Werewolf AI with theory of mind and MCTS minimizes suspicion, wins 49.8%","Reading faces and voices: AI Werewolf agent dodges suspicion in One Night"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result leans on the learned suspicion model being accurate for opponents it was not trained on; if its predicted beliefs do not match what those opponents actually think, the search is planning against a false model and the win-rate gains would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["AI Werewolf agent reads faces and tones to dodge suspicion","Multimodal AI Werewolf agent predicts suspicion, wins 49.8% of games","Faces, tones, and theory of mind: AI Werewolf agent plans around suspicion","Werewolf AI with theory of mind and MCTS minimizes suspicion, wins 49.8%","Reading faces and voices: AI Werewolf agent dodges suspicion in One Night"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000865,"raw_usage":{"total_tokens":3766,"prompt_tokens":977,"completion_tokens":2789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":2682}},"tokens_in":593,"tokens_out":2789,"duration_ms":18895,"temperature":1.0,"reasoning_tokens":2682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:26:55.031473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set MultiMind against a held-out set of opponent agents, record each opponent's true suspicion after every statement, and compare it with the theory-of-mind model's predicted belief matrices; if the predictions are no better than chance, or if the search-chosen utterances do not reduce the opponents' measured suspicion below randomly chosen utterances, the paper's central explanation collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the 5-player One Night Ultimate Werewolf environment and the LLM-instructed and RL-instructed baselines that MultiMind is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the facial-expression classifier that turns video frames into eight emotion labels for human opponents."}],"review_version":1}