AgentRoboCritic - Failure Mode Accuracy: leaderboard
AgentRoboCritic (arXiv:2611.26613v1, Table 1): zero-shot multimodal action-level critic accuracy (%) on classifying the failure mode across 264 human-annotated failed robot-control actions (off-the-shelf models only); higher is better.
Metric: Failure Mode Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 21 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 80.68 | #10 |
| 2 | Claude Fable 5.2 | 72.73 | #1 |
| 3 | GPT-6.1 Sol | 62.12 | #15 |
| 4 | GPT-6.1 Terra | 61.36 | #25 |
| 5 | Claude Opus 5.5 | 60.98 | #9 |
| 6 | GLM-5.3 Flash | 42.8 | #133 |
| 7 | DeepSeek V4.1 Flash (1017) | 32.95 | #47 |
| 8 | Claude Sonnet 5.5 | 32.58 | #24 |
| 9 | Muse Glimmer 30B | 32.58 | #398 |
| 10 | Kimi K3.1 Pro | 32.2 | #36 |
No result here: #3 Gemini 4 Neon, #4 GPT-6.2 Astra, #7 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=agentrobocritic-failure-mode-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-21.