VIABLE - Single-Failure Diagnosis: leaderboard
VIABLE (Zhao et al., 2026) evaluates vision-language models as judges of answers to video questions from visually impaired users, drawn from outdoor walking, indoor scene and hand-object manipulation corpora. Here each answer carries one failure injected by Claude Sonnet 4.6 from a 12-type taxonomy (perception, cognition, action and interaction failures such as a safety violation or an omitted piece of evidence), and the judge must name it: a prediction counts when the injected type is among the judge's two predicted codes. Accuracy in percent over 100,050 answers; higher is better.
Metric: Accuracy (%; 100,050 single-injected answers, the injected failure among the judge's two predicted codes, pooled over the WAD, VisAssist and VIA-EgoDex corpora). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 52.6 | #121 |
| 2 | Claude Sonnet 4.6 | 45.9 | #158 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=viable-single-failure-diagnosis · How It Works · Data refreshed daily, snapshot 2026-11-12.