CollabBench Cook - Trustfulness: leaderboard
Affective collaboration quality in the Cook-MultiPlayer cooking game of CollabBench (2026): trustfulness (reliable commitments and consistent, honest communication) of the evaluated agent, scored by a penalty-based LLM judge over windows of its reasoning, messages and actions, out of 5. Higher is better. The evaluated model plays one seat of a two-player game with a simulated partner whose personality varies.
Metric: Trustfulness score (out of 5): trustfulness (reliable commitments and consistent, honest communication), rated by a penalty-based DeepSeek-V3.1 judge: each window of three consecutive agent actions starts at 5 and loses 1 to 3 points per violation, and the trajectory score subtracts violation and worst-window terms and a message-sparsity penalty, clipped at 0; Cook-MultiPlayer (Overcooked-style cooperative cooking, five kitchen scenarios): the evaluated model drives Agent 1 inside the ProAgent framework while a DeepSeek-V3.1 simulator role-plays a partner with a Big Five personality profile; mean over all evaluation trajectories; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.1 | 2.97 | #497 |
| 2 | GPT-5.2 | 2.89 | #176 |
| 3 | Qwen 2.5 72B Instruct | 2.77 | #1092 |
| 4 | Qwen 2.5 7B Instruct | 1.92 | #2564 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=collabbench-cook-trustfulness · How It Works · Data refreshed daily, snapshot 2026-11-13.