op02/mafia_persona

MafiaPersona

LLM agents with layered persona prompts play seven-player Mafia. The benchmark asks whether conditioned agents manage their personas once told that expressing a trait is costly. This repository holds every game, statement and reasoning trace, the per-agent Empath profiles over all 194 categories, and the statistics reported in the paper.

Composition

475 games across five model families and four conditions (25 games each; Grok 3 Mini has no T/B-only games): 6,775 public statements (531,848 words) and 2,445 reasoning traces.

model_name	games	statements	words	traces
Claude Haiku 4.5	100	1429	130596	0
GPT-5.4 Mini	100	1473	123893	1401
Grok 3 Mini	75	1044	97136	1044
Llama 4 Maverick	100	1480	105538	0
Qwen 2.5 72B Instruct	100	1349	74685	0

Conditions: persona (Layers T, B and G), tb_only (Layer G, the stated cost, removed), baseline (no persona, same seeds), noise_baseline (no persona, different seeds; used for the placebo comparison).

Persona	Role	Taxonomy
Machiavellianism	mafia	Dark Triad
High Neuroticism	detective	Big Five
High Extraversion	doctor	Big Five
Low Agreeableness	citizen	Big Five
Low Conscientiousness	citizen	Big Five

Measurement

Empath profiles: all public statements of an agent in a game are joined, tokenized on whitespace (Empath's default), and scored as rates per 100 tokens for all 194 categories. No agent is excluded. Reasoning traces come from a separate call with reasoning enabled and are never fed back into the game.

Tables

statements (6,775 rows)

column	type	description
statement_id	object	Unique id of the statement
game_key	object	Unique game id:
model_family	object	openai, xai, anthropic, meta or qwen
condition	object	persona, tb_only, baseline or noise_baseline
seed	int64	Game seed (92-116; noise baseline 192-216)
day	int64	Game day
turn_index	int64	Order of the statement within the game
player_id	int64	Seat 0-6
role	object	mafia, detective, doctor or citizen
persona	object	Persona carried by the agent, or none
text	object	Public statement, verbatim
word_count	int64	Whitespace tokens

games (475 rows)

column	type	description
game_key	object	Unique game id:
model_family	object	openai, xai, anthropic, meta or qwen
model_name	object	Model display name
model_id	object	OpenRouter model identifier
condition	object	persona, tb_only, baseline or noise_baseline
seed	int64	Game seed (92-116; noise baseline 192-216)
run_batch	object	main or rebuttal_tb_only (T/B-only games were run later with the same engine)
winner	object	Winning side
game_length	int64	Days played
n_statements	int64	Public statements in the game
config_fingerprint	object	SHA-256 of the run configuration, including persona texts
roles_json	object	Seat -> role
personas_json	object	Seat -> persona
votes_per_day_json	object	Day -> {voter: target}
eliminations_json	object	Seat -> [cause, day]
detective_results_json	object	Detective's investigation results

agents (3,325 rows)

column	type	description
agent_id	object	Unique id of the agent in a game
game_key	object	Unique game id:
model_family	object	openai, xai, anthropic, meta or qwen
condition	object	persona, tb_only, baseline or noise_baseline
seed	int64	Game seed (92-116; noise baseline 192-216)
player_id	int64	Seat 0-6
role	object	mafia, detective, doctor or citizen
persona	object	Persona carried by the agent, or none
persona_construct	object	Construct named in Layer T
persona_taxonomy	object	Dark Triad or Big Five
n_statements	int64	Public statements made by the agent
total_words	int64	Whitespace tokens spoken by the agent
exit	object	DAY_VOTE, MAFIA_KILL or survived
exit_day	float64	Day of elimination
votes_received	int64	Day votes cast against the agent, all days
rounds_survived	float64	Day of a day-vote elimination; that day + 0.5 for a night kill; game length + 1 if the agent survived

traces (2,445 rows)

column	type	description
trace_id	object	Unique id of the trace
game_key	object	Unique game id:
model_family	object	openai, xai, anthropic, meta or qwen
condition	object	persona, tb_only, baseline or noise_baseline
seed	int64	Game seed (92-116; noise baseline 192-216)
day	int64	Game day
player_id	int64	Seat 0-6
role	object	mafia, detective, doctor or citizen
persona	object	Persona carried by the agent, or none
reasoning_text	object	Reasoning trace from a separate call with reasoning enabled; never shown to other players
reasoning_words	int64	Whitespace tokens in the trace
statement_id	object	Unique id of the statement

empath_profiles (3,325 rows)

column	type	description
agent_id	object	Unique id of the agent in a game
game_key	object	Unique game id:
model_family	object	openai, xai, anthropic, meta or qwen
condition	object	persona, tb_only, baseline or noise_baseline
seed	int64	Game seed (92-116; noise baseline 192-216)
player_id	int64	Seat 0-6
role	object	mafia, detective, doctor or citizen
persona	object	Persona carried by the agent, or none
n_tokens	int64	Whitespace tokens scored
194 Empath categories (achievement ... zest)	float	Rate per 100 whitespace tokens

judge_ratings (10,284 rows)

One row per judge call. Each of three judges, from different model families, rated each agent's full set of public statements twice, blind to condition.

column	type	description
agent_id	object	Agent in a game; joins to agents
game_key	object	Unique game id:
model_family	object	openai, xai, anthropic, meta or qwen
condition	object	persona or baseline
seed	int64	Game seed
player_id	int64	Seat 0-6
role	object	mafia, detective, doctor or citizen
persona	object	Persona carried by the agent, or none
word_count	int64	Whitespace tokens the judge read
judge_id	object	judge_1_openai, judge_2_google or judge_3_anthropic
judge_model	object	openai/gpt-4o-mini, google/gemini-2.5-flash or anthropic/claude-haiku-4-5
ordering	object	standard or reversed presentation; each judge rated each agent once in each
dominance, suspicion, analytical, anxiety, sociability	int64	Rating on a 1-5 scale

judge_panel (1,714 rows)

One row per rated agent: the panel score, the mean over three judges and both orderings.

column	type	description
agent_id ... word_count		As in judge_ratings
dominance, suspicion, analytical, anxiety, sociability	float64	Mean of six ratings, 1-5

effects (8,730 rows)

Every persona x category x model comparison (4,850 rows) and every cell of the placebo comparison, baseline vs noise baseline within role (3,880 rows). Each comparison uses one observation per game: the mean profile of the agents carrying the persona (or holding the role), against the baseline agents in the same role, 25 games per arm. Every row is reproducible from empath_profiles.

column	type	description
analysis	object	persona_vs_baseline or placebo_baseline_vs_noise
model_family	object	openai, xai, anthropic, meta or qwen
unit	object	Persona (persona_vs_baseline) or role (placebo)
category	object	Empath category (all 194)
delta	float64	Difference in mean rate per 100 tokens
d	float64	Cohen's d, pooled variance
ci_lo, ci_hi	float64	95% bootstrap interval for d, 10,000 resamples of games (seed 0)
p_perm	float64	Two-sided permutation p, 10,000 permutations, (1 + hits) / 10,001
bh	bool	Significant after Benjamini-Hochberg at q = 0.05 over 194 categories per unit and model
testable	bool	False when the category never occurs in either arm (d undefined)

preregistration (25 rows)

The 25 directional predictions deposited on OSF before any game was run, with how many model-level tests were testable and how many matched.

column	type	description
prediction_id	int64	1-25
persona	object	Persona
category	object	Predicted Empath category; dominant, independent and sympathetic are not Empath categories and are untestable
predicted_direction	object	+ (increase) or - (decrease) relative to baseline
n_testable, n_match	int64	Model-level tests that were testable, and that matched

preregistration_tests (125 rows)

One row per prediction x model. A category with zero occurrences for a model and persona is untestable (a priori rule).

column	type	description
prediction_id ... predicted_direction		As in preregistration
model_family	object	openai, xai, anthropic, meta or qwen
testable	bool	Testable under the a priori rule
observed_d	float64	Observed Cohen's d (from effects)
matches_prediction	object	Observed direction matches the prediction

raw/<family>/<condition>/seed_<n>.json holds the original game logs.

Intended use and license

Research on persona conditioning, multi-agent interaction, and monitoring of model reasoning and outputs. All content is synthetic, generated by LLM agents in a simulated game, and contains no personal information. Some personas (e.g. Machiavellianism) produce deceptive or confrontational in-game dialogue by design. Released under CC-BY-4.0.