FailBench - reflect: leaderboard
FailBench evaluates video-language models and specialist detectors judging robot task success across real and simulated robot videos under a shared prompt, input and decoding recipe. Macro results average two-class datasets; the failure-only reflect subset uses failure recall. This series reports FailBench - reflect from 2609.03611v1, S4.T2: Failure Recall (%). Higher is better.
Metric: Failure Recall (%). Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 31B (IT) | 73 | #180 |
| 2 | Qwen 3 VL 8B (Thinking) | 57 | #423 |
| 3 | GPT-4o | 53 | #358 |
| 4 | Gemini 3 Flash | 40 | #105 |
| 5 | Qwen 3 VL 4B (Thinking) | 33 | #546 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=failbench-reflect · How It Works · Data refreshed daily, snapshot 2026-11-07.