CAME-Bench - Small: 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 small split (six trajectories averaging 23K tokens, 144 questions), answered with the whole trajectory in context.
Metric: Answer-set macro F1 (%; 144 questions on six trajectories averaging 23K tokens; full-context answer, gpt-4.1-mini judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini (Medium) | 80.4 | #349 (GPT-5 Mini) |
| 2 | GPT-4.1 Mini | 71.2 | #775 |
| 3 | Qwen 3 235B A22B 2507 Instruct | 70.4 | #619 |
| 4 | GPT-4o Mini | 31.5 | #1627 |
| 5 | DeepSeek V3.1 | 22.8 | #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-small · How It Works · Data refreshed daily, snapshot 2026-11-12.