MP-Bench (Failure Attribution) - Hand-Crafted MAS Reasoning: leaderboard
Metric: Attribution-reasoning score on a 1-10 scale: GPT-5.1 (temperature 0) rates the consolidated failure reasons and ideal actions against the GPT-5.1-consolidated annotator rationales, averaged over failure-inducing steps, on MP-Bench's 169 failed hand-crafted (Magentic-One) multi-agent executions from GAIA and AssistantBench, three expert annotators per execution log; the model runs All-at-Once attribution zero-shot 3 times at temperature 1.0 and its runs are consolidated by consensus rate; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 7.95 | #138 |
| 2 | GPT-4.1 | 7.63 | #240 |
| 3 | GPT-5.1 | 7.62 | #131 |
| 4 | GPT-OSS-120B | 7.51 | #330 |
| 5 | O3 Mini | 7.48 | #266 |
| 6 | Qwen 3 8B | 6.11 | #667 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mp-bench-failure-attribution-hand-crafted-mas-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-11.