M3-Bench - WW: leaderboard
Evaluates LLM agents' social behaviour in mixed-motive games on M3-Bench (Xie et al., 2026): 24 game-theoretic tasks across four progressive levels, each played against a fixed opponent pool with and without free-form messaging. Scores are the BTA (Behavioral Trajectory Analysis) view: rule-based indicators computed from each agent's logged actions and payoffs, normalised within the task and reported on a 0-100 scale. The scale is normative, so cooperation, fairness and joint payoff efficiency raise it and exploitation lowers it. WW is the mini Werewolf variant, a Level 4 social-deduction game of hidden roles, deception and alliance shift. The paper's RPA (reasoning) and CCA (communication) views are LLM-judge scores of the same episodes and are not included. This series reports M3-Bench - WW using BTA Score.
Metric: BTA Score. Source: huggingface.co. Saturation forecast: Around March 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 83 | #87 |
| 2 | Kimi K2 (Thinking) | 82 | #252 (Kimi K2) |
| 3 | GPT-5.1 (Non-reasoning) | 80 | #145 (GPT-5.1) |
| 4 | GPT-5 Pro | 76 | #89 |
| 5 | GPT-OSS-120B | 75 | #346 |
| 6 | Mistral Large 3 | 70 | #410 |
| 7 | DeepSeek V3.2 (Non-reasoning) | 68 | #210 (DeepSeek V3.2) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=m3-bench-ww · How It Works · Data refreshed daily, snapshot 2026-11-05.