RoboAbstention - False Premise: leaderboard
Metric: Abstention rate (%) on the 957 False Premise 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 2041. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 28.5 |
| 2 | Llama 4 Maverick | 15.5 |
| 3 | Qwen 3.5 27B | 7.6 |
| 4 | Claude Sonnet 4.6 | 7.2 |
| 5 | GPT-5.4 | 3.6 |
| 6 | GPT-5.4 Mini (High) | 3.4 |
| 7 | GPT-5.4 Mini (Low) | 3.1 |
| 8 | GPT-5.4 Mini (Non-reasoning) | 3 |
| 9 | GPT-5.4 Mini (Medium) | 2.9 |
| 10 | GPT-5.4 Nano | 1.5 |
Interactive version: theaggregate.ai/benchmark?slug=roboabstention-false-premise · How It Works · Data refreshed daily, snapshot 2026-10-07.