RoboAbstention - Contradictory Instructions: leaderboard
Metric: Abstention rate (%) on the 837 Contradictory Instructions instructions, image plus instruction to an embodied VLM planner, every instruction warranting abstention, act or abstain judged by GPT-5.4 mini (97.5 percent agreement with human labels), API default sampling; higher is better. Source: arxiv.org. Saturation forecast: Around 2035. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 36.3 |
| 2 | GPT-5.4 | 29 |
| 3 | Llama 4 Maverick | 28.6 |
| 4 | Claude Sonnet 4.6 | 26.5 |
| 5 | GPT-5.4 Nano | 22.6 |
| 6 | GPT-5.4 Mini (Low) | 22 |
| 7 | GPT-5.4 Mini (Medium) | 20.5 |
| 8 | GPT-5.4 Mini (High) | 17.2 |
| 9 | GPT-5.4 Mini (Non-reasoning) | 16.8 |
| 10 | Qwen 3.5 27B | 5.4 |
Interactive version: theaggregate.ai/benchmark?slug=roboabstention-contradictory-instructions · How It Works · Data refreshed daily, snapshot 2026-10-07.