Trust, Lies, and Long Memories: Emergent Social Dynamics and Reputation in Multi-Round Avalon with LLM Agents
arXiv:2604.20582v1 [cs.MA] 22 Apr 2026
Suveen Ellawela [email protected] National University of Singapore https://github.com/SuveenE/multi-round-avalon-agents
Figure 1: Overview of our experimental setup showing emergent social dynamics and reputation. Five LLM agents (Alice, Bob, Charlie, Diana, Eve) play repeated Avalon games with randomized roles. After each game, agents generate self-reflections and observations about other players, which persist into subsequent games as memory. This cross-game memory enables reputation formation, with agents referencing past behavior in their strategic reasoning (example quotes shown on right). Abstract. We study emergent social dynamics in LLM agents playing The Resistance: Avalon, a hidden-role deception game. Unlike prior work on single-game performance, our agents play repeated games while retaining memory of previous interactions (who played which roles and how they behaved), enabling us to study how social dynamics evolve. Across 188 games, two key phenomena emerge. First, reputation dynamics emerge organically when agents retain cross-game memory: agents reference past behavior in statements like “I’m wary of repeating last game’s mistake of over-trusting early success.” These reputations are role-conditional: the same agent is described as “straightforward” when playing good but “subtle” when playing evil, and high-reputation players receive 46% more team inclusions. Second, higher reasoning effort supports more strategic deception: evil players more often pass early missions to build trust before sabotaging later ones (75% in high-effort games vs 36% in low-effort games). Together, these findings show that repeated interaction with memory gives rise to measurable reputation and deception dynamics among LLM agents. Keywords: Large Language Models, Multi-Agent Systems, Multi-Round Interactions, Deception Games, Social Dynamics, Theory of Mind
1
1
Introduction
“aggressive when evil” but “cautious when good.” First, we introduce cross-game memory to study how such reputation and trust dynamics develop over repeated interactions between the same LLM agents. Second, we systematically vary reasoning depth to isolate its effect on strategies used during game play. Third, we conduct detailed qualitative analysis of the natural language strategies that emerge, cataloging role-conditional reputation patterns and behavioral “tell” taxonomies that agents use to identify hidden roles.
Social deduction games provide a compelling testbed for studying emergent behavior in AI systems. Unlike chess or Go, where optimal play can be computed, games like The Resistance: Avalon [Wikipedia, 2024] require social reasoning: inferring hidden information from behavior, building trust, detecting deception, and coordinating without explicit communication channels. These capabilities are central to real-world multi-agent scenarios, from negotiation to collaborative problem-solving. Recent advances in large language models (LLMs) have enabled agents that can engage in extended natural language dialogue, maintain context across interactions, and reason about others’ mental states [Kosinski, 2024]. This raises fundamental questions about multi-agent multi-round dynamics: Do LLMs develop stable social models when they interact repeatedly with the same agents across multiple games? Are these models role-conditional, where agents distinguish how the same individual behaves as deceiver versus cooperator? We address these questions using The Resistance: Avalon, a hidden-role game where a “good” team attempts to complete missions while “evil” players sabotage secretly. The game includes an assassination mechanic: if the good team wins, evil can still claim victory by identifying “Merlin,” a good player who knows evil’s identities but must hide this knowledge. Our contributions are as follows: • We show that LLM agents develop stable, roleconditional reputations when given cross-game memory: the same player is described as “subtle” when evil but “straightforward” when good. Agents actively reference previous games and their learnings when making decisions. • We find that high-reputation players receive 45.6% more team inclusions, showing that reputation has downstream strategic consequences for coalition formation. • We find that higher reasoning enables more sophisticated deception: evil players more frequently pass early missions to build trust before sabotaging, a strategy known as “sleeper agent” among human players (75% at medium/high reasoning vs 36% at low reasoning).
2.2
LLMs in Strategic Multi-Agent Settings
The broader literature on LLMs in strategic settings has examined negotiation, where Lewis et al. [2017] trained end-toend dialogue agents for deal-making, and diplomacy, where Meta’s Cicero [Meta Fundamental AI Research Diplomacy Team (FAIR), 2022] achieved human-level play by combining language models with strategic planning. Park et al. [2023] demonstrated that LLM agents in simulated social environments develop emergent behaviors like information diffusion and relationship formation. The AgentBench framework [Liu et al., 2023] provides comprehensive evaluation of LLM agents across code generation, web browsing, and other single-agent environments, establishing that frontier models substantially outperform smaller models on agentic tasks. However, multi-agent social deduction, where agents must reason about hidden information, form coalitions, and detect deception through behavioral analysis, remains underexplored. Our work addresses this gap by studying repeated games with persistent memory, enabling investigation of how agents build, maintain, and exploit social models over time.
2.3
Deception and Theory of Mind
Detection of deceptive behavior in AI systems is an active area of AI safety research. Hubinger et al. [2024] demonstrated that deceptive behaviors can persist through safety training, raising concerns about detecting misaligned AI systems. Avalon provides a complementary perspective: a controlled setting where ground truth (role assignments) is known, enabling precise measurement of detection accuracy and systematic analysis of failure modes. On the cognitive side, Kosinski [2023] found evidence that theory of mind capabilities may emerge in large language models, with GPT2 Related Work 4 passing classic false-belief tasks. Our work extends this to dynamic social contexts where mental models must be 2.1 LLMs in Social Deduction and Board continuously updated based on behavioral evidence across Games extended multi-party interactions, a substantially more deMost directly related to our work is AvalonBench [Light manding test of social cognition than static vignette-based et al., 2023], which introduced a benchmark for evaluat- assessments. ing LLM agents in The Resistance: Avalon. AvalonBench focused on single-game performance with rule-based oppoThe Resisnents; our work extends this in three key directions, motivated 3 Game Environment: by bringing the game closer to how humans actually play it tance: Avalon through multi-agent multi-round dynamics. When humans play repeatedly with the same group, they develop mental This section provides a detailed description of The Resistance: models of how each player behaves differently depending Avalon, the hidden-role social deduction game used in our on their team assignment, noticing that someone might be experiments. 2
Figure 2: Game flow in The Resistance: Avalon. Each mission round consists of four phases: (1) Discussion, where players share observations and suspicions; (2) Team Proposal, where the current leader selects team members; (3) Voting, where all players approve or reject the proposal; and (4) Mission Execution, where approved team members secretly choose success or fail. The cycle repeats until one team wins three missions.
3.1
3.3
Overview and Objectives
The Resistance: Avalon is a game of hidden loyalties for 5–10 players. Players are secretly divided into two teams: the Good Team (Loyal Servants of Arthur), who form the majority and aim to successfully complete three of five missions, and the Evil Team (Minions of Mordred), a minority who know each other’s identities and aim to either sabotage three missions or identify and assassinate Merlin. The game captures a fundamental tension: good players must coordinate without certain knowledge of who to trust, while evil players must deceive while appearing trustworthy.
Team Composition
The ratio of good to evil players varies with player count. In the 5-player variant used in most of our experiments, the Good Team consists of Merlin and 2 Loyal Servants, while the Evil Team consists of the Assassin and 1 generic Minion. Role configurations for other player counts (6–10 players), which introduce additional special roles like Percival, Morgana, Mordred, and Oberon, are provided in Section B.
3.4
Game Flow
A game of Avalon proceeds through the following phases:
3.2
Special Roles
Beyond basic good/evil alignment, Avalon includes special Phase 1: Role Assignment and Night Phase At game start, each player is secretly assigned a role. During the roles with unique information or abilities: “night phase,” players receive their private information: evil players (except Oberon) learn each other’s identities, Merlin Table 1: Special roles in The Resistance: Avalon. learns which players are evil (except Mordred), and Percival learns which players appear as Merlin (the true Merlin plus Role Team Ability Morgana, without knowing which is which). This informaMerlin Good Knows evil; must guide tion asymmetry is the foundation of all strategic interaction. subtly Percival Assassin Morgana Mordred Oberon
Good Evil Evil Evil Evil
Sees Merlin & Morgana Can assassinate Merlin Appears as Merlin Hidden from Merlin Isolated from evil team
Phase 2: Mission Rounds (Repeated up to 5 times) The game consists of up to 5 mission rounds. Each round proceeds through four steps: 1. Team Proposal: The current Leader proposes a team of The interplay of these roles creates rich strategic comspecified size for the mission and explains their reasonplexity. Merlin must guide without revealing; Percival must ing. protect without certainty; evil must deceive while coordinat2. Discussion: All players debate the proposed team, arguing. ing for or against inclusion of specific individuals. This 3
is the primary opportunity for information exchange and deception. 3. Team Vote: All players simultaneously vote to Approve or Reject; if a majority approves, the team proceeds to the mission, while rejection passes leadership clockwise. After 4 consecutive rejections, the 5th proposal autoapproves. 4. Mission Execution: Approved team members secretly choose Success or Fail. Good players must choose Success, while evil players may choose either strategically. Any Fail card fails the mission, except Mission 4 in 7+ player games which requires 2 Fails. In a 5-player game, missions require teams of 2, 3, 2, 3, and 3 players respectively. Full mission team sizes for all player counts are provided in Section B.
receive their last 3 self-assessments along with accumulated observations about each player. This enables longitudinal study of reputation formation while keeping context manageable. Note that reputation in our setting reflects both behavioral style and known alignment history.
4.3
We manipulate reasoning depth using the reasoning_effort parameter available in GPT-5.1, across three levels: • Low: Minimal extended thinking • Medium: Moderate extended thinking • High: Maximum extended thinking This parameter controls the token budget allocated to internal reasoning before generating a response. Higher reasoning effort results in longer internal deliberation and typically more structured analysis. Average thinking time per decision was 7.5 seconds for Low, 37.5 seconds for Medium, and 107 seconds for High.
Phase 3: Victory Determination The game ends when either team achieves victory. Good wins by successfully completing 3 of 5 missions. Evil wins either by failing 3 missions or through assassination: if good wins by missions, the Assassin gets one chance to identify Merlin, and a correct guess gives evil the victory instead.
4
4.4
Table 2: Dataset overview. Dataset
Agent Architecture
Following the ReAct paradigm [Yao et al., 2023], each agent is instantiated as a prompted LLM (OpenAI GPT-5.1 [OpenAI, 2025]) that interleaves reasoning with action. Each agent receives: • Role knowledge: Private information about their role and, for some roles, others’ identities • Game state: Public mission history, vote records, and discussion transcripts • Memory (tournament mode): Reflections from previous games including observations about other players Agents generate natural language discussion contributions, vote on team proposals, and (if on approved mission teams) choose to succeed or fail the mission.
4.2
Experimental Conditions
We collected 188 games organized into four disjoint datasets, summarized in Table 2.
Methods
Our experimental framework enables LLM agents to play repeated Avalon games while retaining memory of past interactions. This section describes the agent architecture, memory system, reasoning manipulation, and data collection procedures.
4.1
Reasoning Effort Manipulation
Games
N
Mem
Reas.
A: Reputation B: Count (Mem) C: Count (No Mem) D: Reasoning
50 60 60 18
5 5–10 5–10 5
Full Full None Full
Low Low Low L/M/H
Total
188
Dataset A enables deep analysis of reputation formation over 50 repeated games with the same 5 agents. Datasets B and C form matched pairs for isolating memory effects across player counts. Dataset D varies reasoning depth to explore how computational budget affects agent behavior.
4.5
Text Analysis Methods
Our qualitative analyses rely on systematic extraction from agent-generated text: Descriptor frequency (Section 5.1): We compiled a seed list of 15 behavioral descriptors commonly used in social evaluation (straightforward, subtle, cautious, trustworthy, quiet, aggressive, reliable, suspicious, etc.) and counted exactmatch occurrences in post-game reflection texts.
Memory System
In tournament mode, agents retain memory across games through a structured reflection system. At the end of each game, all roles are revealed to all agents (mirroring standard human play). Agents then generate a post-game reflection containing a self-assessment of their own performance and observations about each other player’s behavior; these reflections may therefore include explicit role identities (e.g., “Bob was evil this game”). At the start of subsequent games, agents
Cross-game references (Section 5.1): We identified references using keyword patterns indicating temporal continuity: “past games,” “last game,” “usually,” “tends to,” “historically,” “track record,” and “previous.” 4
4.6
Implementation Details
Game 3 (Bob): “I’d slightly prefer an Alice + Diana pair to start, since both tend to play pretty straightforwardly early.”
Discussion proceeded in fixed turn order starting from the leader, with one message per player per proposal round, softly capped at 2 sentences via prompting. Roles were sampled uniformly at random for each game, subject to Avalon role constraints for the given player count (e.g., exactly one Merlin, one Assassin, etc.). Player names (Alice, Bob, Charlie, Diana, Eve, etc.) were fixed across all games within a dataset. All discussions, player reflections, and memories were saved for all games, enabling detailed post-hoc analysis. All prompts are provided in Section A.
5
Game 23 (Eve): “I slightly prefer Alice + Bob over Alice + Diana for Mission 1. First mission failing puts us in a hole fast, so I’d rather start with the pair that historically plays a bit more conservatively.” 5.1.3
Critically, these descriptions are role-conditional. Because roles are revealed at the end of each game, post-game reflections capture how agents retrospectively explain behavior conditional on ground truth. Table 4 shows how “subtle” and “straightforward” descriptors vary by the target’s actual role.
Results
We organize our findings around three main phenomena: the emergence of reputation dynamics when agents retain cross-game memory, the downstream effects of reputation on strategic behavior, and the relationship between reasoning depth and strategic behavior.
Table 4: Descriptor usage by target’s actual role. “Subtle”
5.1 Reputation Dynamics: Emergence of Stable, Role-Conditional Models
“Straightforward”
Player
Evil
Good
Evil
Good
Bob Eve Alice
16 16 3
12 9 6
0 1 1
27 16 28
Key finding: Players are described as “straightforward” dramatically more often when playing good roles than evil. Bob receives this descriptor 27 times when good but zero times when evil; Eve shows 16 vs 1; Alice shows 28 vs 1. This demonstrates that the same player’s behavior is perceived systematically differently depending on their role.
Our primary question is whether LLM agents develop stable social models when interacting repeatedly with the same individuals. To investigate this, we analyze Dataset A, where five agents played 50 consecutive games while retaining memory of previous interactions. 5.1.1
Role-Conditional Descriptions
Descriptor Convergence
5.2 Reputation Effects on Coalition Formation
We first examine whether agents develop consistent perceptions of each other over time. After each game, agents generate reflections describing other players’ behavior. We analyzed these reflections for recurring behavioral descriptors (Table 3).
The emergence of reputation raises a natural question: does it actually influence strategic behavior? In Avalon, one of the most consequential decisions is team selection: leaders propose teams for each mission, and being included on teams is essential for both gathering information and influencing Table 3: Descriptor frequency by player (50-game tourna- outcomes. We examined whether players with stronger reputations received more team invitations. ment). To measure reputation, we counted positive descriptors (trustworthy, straightforward, solid, safe, reliable, etc.) in Player #1 (count) #2 (count) #3 (count) post-game reflections. At game 20, cumulative counts were: Alice straightforward (29) cautious (25) trustworthy (21) Alice (76), Diana (63), Charlie (49), Bob (42), Eve (29). We Bob subtle (28) straightforward (27) cautious (26) classified the top 2 players (Alice, Diana) as high-reputation Charlie subtle (38) straightforward (25) cautious (23) Diana subtle (35) straightforward (25) quiet (16) and bottom 2 (Bob, Eve) as low-reputation, excluding the Eve cautious (26) subtle (25) quiet (25) middle player. We then counted team inclusions on approved missions for games 21–50. Charlie receives “subtle” 38 times, significantly more than any other player, establishing a consistent reputation that Table 5: Team inclusion by reputation tier (Games 21–50). persists across games. Reputation Tier Total Inclusions Avg per Game 5.1.2 Cross-Game Behavioral References High (top 2 players) 150 4.84 Low (bottom 2 players) 103 3.32 Beyond forming stable impressions, do agents actively use their memories when making decisions? We searched for explicit references to past games in discussion transcripts and Effect size: +45.6% more inclusions for high-reputation found 105 instances where agents cited historical behavior to players. This correlation suggests that emergent reputation justify their positions: has downstream strategic consequences. We verified that the 5
effect holds when excluding self-inclusions: high-reputation The pattern shows: (1) initial discovery of cross-game players still received 38% more inclusions from others. exploitation, (2) peak meta-awareness with explicit warnings, (3) normalization as anti-anchoring becomes standard practice.
5.3
Reasoning Depth and Strategic Behavior
Dataset D varies reasoning effort across 18 five-player games to explore how computational budget affects agent behavior. 6 Discussion We discovered that higher reasoning correlates with more sophisticated evil team strategies. Our experiments reveal that LLM agents, when given the ability to remember past interactions, develop social dynamics that mirror aspects of human group behavior. We discuss the 5.3.1 Trust-Building Through Early Cooperation implications of these findings and acknowledge limitations A sophisticated deception strategy in Avalon is for evil play- of our study. ers to pass early missions, building trust and credibility before sabotaging later when the stakes are higher. We found this behavior emerges more frequently at higher reasoning levels 6.1 Implications for AI Social Reasoning (Table 6). The emergence of reputation, coalition preferences, and metaTable 6: Evil players passing early missions by reasoning strategic awareness suggests that LLM agents can develop level (5-player games). sophisticated social models through experience. Three capabilities are particularly notable: Level Games Pass Early % 1. Role-conditional modeling: Agents’ retrospective deLow 6 0 0% scriptions of the same individual differ systematically Medium 6 5 83% depending on that player’s true alignment, suggesting High 6 4 67% they learn to recognize distinct behavioral signatures for good versus evil play. For comparison, across all other 5-player games with Low 2. Reputation exploitation: Agents leverage built repureasoning (Datasets A, B, C), this strategy appeared in 36% tation for coalition formation, demonstrating strategic of games (27/76). The increase at Medium/High reasoning social cognition. (75%, 9/12 games) suggests that additional computation en3. Meta-adaptation: Agents recognize and adapt to metaables more sophisticated deception timing. level patterns, engaging in an “arms race” of strategy and counter-strategy. Notable example (Medium, Game 3): Eve passed both Mission 1 and Mission 3 before finally sabotaging, a patient approach that built substantial trust before striking. 5.3.2
6.2 Reasoning Depth and Strategic Sophistication
Assassination Accuracy
The emergence of sleeper agent strategies at higher reasoning levels (75% vs 36% at low reasoning) suggests that extended computation enables more sophisticated strategic planning. Rather than simply improving reactive decision-making, additional reasoning budget appears to unlock long-term deceptive strategies that require patience and delayed gratification.
As a secondary observation, we noted that assassination accuracy (evil’s ability to identify Merlin after good wins) also trends upward with reasoning: 67% (Low) → 75% (Medium) → 100% (High). However, the small sample sizes (3–4 attempts per condition) preclude strong conclusions.
5.4
Meta-Strategic Adaptation 6.3
Limitations
As reputations form, a strategic tension emerges: relying on past behavior to predict future actions can be exploited Several limitations constrain our findings: by adversaries who recognize this pattern. We examined • Our reasoning comparison uses only 6 games per conwhether agents develop awareness of this meta-level dynamic dition; larger samples would provide tighter confidence and adapt accordingly. intervals on the assassination accuracy trends. Game 35 (Bob): “I get why you like you+Diana, but anchor• All agents use similar base models from the same family, ing off past games can be a trap if either of you rolled evil and cross-model comparisons would test whether these this time.” social dynamics generalize across architectures. • Agent behavior depends substantially on prompting, and Game 35 (Eve): “I’d rather avoid recycling ‘trusted’ pairs different prompt designs might yield different social from past games too; something like Bob+Charlie gives us a dynamics or reputation patterns. fresh read.” 6
7
Conclusion
Wikipedia. The resistance (game). https://en. wikipedia.org/wiki/The_Resistance_(game), 2024. Accessed: 2025-02-04.
We studied emergent social dynamics in LLM agents playing The Resistance: Avalon across 188 games varying in player count, memory, and reasoning depth. Our findings reveal that: 1. Reputation dynamics emerge organically: Agents develop stable, role-conditional models of each other, describing the same player differently when they play good versus evil roles. Agents actively reference previous games when making decisions. 2. Reputations have strategic consequences: Highreputation players receive 46% more team inclusions. 3. Reasoning depth enables sophisticated deception: Evil players passing early missions to build trust before sabotaging later appears in 75% of higher-reasoning games versus 36% at low reasoning. These findings demonstrate that LLMs can develop nuanced social reasoning capabilities in multi-agent settings, with implications for AI safety, human-AI collaboration, and computational social science.
Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In ICLR, 2023.
A
Agent Prompts
This appendix documents all prompts used in our experiments.
A.1
Role-Specific Knowledge
Each agent receives role-specific context at the start of each game phase. Merlin: You are Merlin. You know these evil players: {evil_list}. Help good win WITHOUT revealing your identity, or the Assassin will kill you!
References
Percival:
Evan Hubinger et al. Sleeper agents: Training deceptive llms that persist through safety training. arXiv preprint arXiv:2401.05566, 2024.
You are Percival (good team). You see these players as Merlin: {merlin_and_morgana}. One is the real Merlin, one might be Morgana (evil). Protect Merlin!
Michal Kosinski. Theory of mind may have spontaneously emerged in large language models. arXiv preprint Assassin: arXiv:2302.02083, 2023.
You are the Assassin (evil team). Your evil teammates
Michal Kosinski. Evaluating large language models in theory are: {evil_teammates}. Sabotage missions. If good wins of mind tasks. Proceedings of the National Academy of 3 missions, you guess who Merlin is! Sciences, 121(45):e2405460121, 2024. doi: 10.1073/pnas. Morgana: 2405460121. You are Morgana (evil team). Your evil teammates are: {evil_teammates}. You appear as Merlin to Percival. Deceive and sabotage!
Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra. Deal or no deal? end-to-end learning of negotiation dialogues. In EMNLP, 2017.
Jonathan Light, Min Cai, Sheng Shen, and Ziniu Hu. Avalon- Mordred: bench: Evaluating llms playing the game of avalon. arXiv You are Mordred (evil team). Your evil teammates are: preprint arXiv:2310.05036, 2023. {evil_teammates}. You are invisible to Merlin. Sabotage missions!
Xiao Liu, Hao Yu, Hanchen Zhang, et al. Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688, Oberon: 2023.
You are Oberon (evil team). You don’t know who your
Meta Fundamental AI Research Diplomacy Team (FAIR). teammates are, and they don’t know you. Sabotage Human-level play in the game of diplomacy by combining missions and try to identify your team! language models with strategic reasoning. Science, 378 Generic Evil: (6624):1067–1074, 2022. OpenAI. Gpt-5.1. gpt-5-1/, 2025.
https://openai.com/index/ You are on the evil team. Your evil teammates are:
{evil_teammates}. Sabotage missions and deceive the good players!
Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generic Good: Generative agents: Interactive simulacra of human behav- You are on the good team. Deduce who the evil players are and ensure missions succeed! ior. In UIST, 2023. 7
A.2
Discussion Phase Prompt
playstyle or patterns", "PlayerName2": "Brief observation...", ...
It’s your turn to speak. Provide a strategic comment about who to trust or who should be on the mission team. Be natural and conversational. Keep it to 1-2 sentences.
} } Make observations about ALL other players (not yourself).
For evil players, an additional instruction is appended: “Remember to deceive and create confusion while appearing trustworthy.” For Merlin: “Subtly guide the team without revealing you know who the evil players are.”
A.3
B Game Configuration by Player Count
Team Proposal Prompt
5 players: Merlin, 2 Loyal Servants (Good) vs Assassin, 1 Minion (Evil). 6 players: Merlin, Percival, 2 Loyal Servants (Good) vs Morgana, Mordred (Evil). Mordred performs assassination. 7 players: Merlin, Percival, 2 Loyal Servants (Good) vs Morgana, Mordred, Oberon (Evil). Morgana performs assassination. 8 players: Merlin, Percival, 3 Loyal Servants (Good) vs Morgana, Mordred, Assassin (Evil). 9 players: Merlin, Percival, 4 Loyal Servants (Good) vs Morgana, Mordred, Assassin (Evil). 10 players: Merlin, Percival, 4 Loyal Servants (Good) vs Morgana, Mordred, Oberon, Assassin (Evil).
You are the mission leader. Propose a team of {size} players for this mission. Available players: {player_list} Respond ONLY with a JSON object: {"team": ["Name1", "Name2", ...], "reasoning": "why you chose this team"}
A.4
Vote Prompt
Vote on this team proposal. Respond ONLY with JSON: {"vote": "approve" or "reject", "comment": "brief reason"}
A.5
Mission Execution Prompt (Evil Only)
Table 7: Mission team sizes by player count. * = requires 2 Good players automatically play Success. Evil players re- Fails. ceive: You’re on the mission. As an evil player, choose ’success’ or ’fail’ strategically. Respond with JSON: {"action": "success" or "fail", "reasoning": "why"}
A.6
C
Discuss who you think Merlin is among the good players. Analyze their behavior and statements in first person (as yourself). Be specific and analytical. Keep it to 2-3 sentences. Speak naturally as if talking to your evil teammates.
6p
7p
8p
9p
10p
2 3 2 3 3
2 3 4 3 4
2 3 3 4* 4
3 4 4 5* 5
3 4 4 5* 5
3 4 4 5* 5
Dataset Statistics Table 8: Tournament results by player count.
Assassination Decision Prompt
Based on all the discussions and your teammates’ analysis, choose who you think is Merlin from the good players. Respond ONLY with JSON: {"guess": "PlayerName", "reasoning": "your analysis in 2-3 sentences"}
A.8
5p
1 2 3 4 5
Evil Team Discussion Prompt
When good wins 3 missions, evil players discuss before assassination:
A.7
M
N
Games
Evil%
Good%
Assn.
5 6 7 8 9 10
10 10 10 10 10 10
60% 50% 60% 100% 50% 80%
40% 50% 40% 0% 50% 20%
20% 29% 20% N/A 0% 33%
Note: The 100% evil win rate at 8 players is an outlier with no clear explanation.
Post-Game Reflection Prompt
Table 9: Memory effect by player count.
After each game, agents generate reflections for cross-game memory: Reflect on your performance in this game. Respond with JSON: { "self_assessment": "What you did well and what you could improve (2-3 sentences)", "player_observations": { "PlayerName1": "Brief observation about their
8
N
Mem
No Mem
Diff
5 6 7 8 9 10
60% 50% 60% 100% 50% 80%
60% 60% 80% 90% 90% 100%
0pp −10pp −20pp +10pp −40pp −20pp
D
Descriptor List
The 15-descriptor seed list: straightforward, subtle, cautious, trustworthy, quiet, aggressive, reliable, suspicious, measured, conservative, transparent, cooperative, deceptive, defensive, strategic.
E
Game Viewer Interface
We developed an interactive web interface to browse through all 188 LLM game plays in our dataset. The viewer (Figure 3) displays player roles, mission progress, and phase navigation, allowing users to explore discussion, proposal, voting, and execution phases.
Figure 3: Game viewer interface showing a 5-player game during Mission 3’s proposal phase.
9