CAME-Bench - Large: leaderboard
CAME-Bench (Yang et al., University of Illinois Urbana-Champaign, 2026) tests memory over long, goal-oriented multi-agent trajectories (travel planning and policy debate) with interleaved goals, reused entities and indirect references, through free-response questions scored by answer-set macro F1 in percent with a gpt-4.1-mini judge; higher is better. This board is the large split (two trajectories averaging 408K tokens, 61 questions), answered with the whole trajectory in context, left-truncated when it exceeds the model's window.
Metric: Answer-set macro F1 (%; 61 questions on two trajectories averaging 408K tokens; full context, left-truncated to the model's window, gpt-4.1-mini judge). Source: arxiv.org. Saturation forecast: Around July 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 Mini | 21.3 | #775 |
| 2 | GPT-5 Mini (Medium) | 21.2 | #349 (GPT-5 Mini) |
| 3 | GPT-4o Mini | 1.7 | #1627 |
| 4 | Qwen 3 235B A22B 2507 Instruct | 1.6 | #619 |
| 5 | DeepSeek V3.1 | 0 | #500 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=came-bench-large · How It Works · Data refreshed daily, snapshot 2026-11-12.