RoboAbstention: leaderboard

RoboAbstention (2026, The Yes-Man Syndrome) gives vision-language robot planners 6,069 image and instruction pairs that all call for abstaining or asking for clarification: missing or ambiguous referents, subjective or underspecified intent, infeasible, contradictory or false-premise requests. The score is the abstention rate in percent; higher is better.

Metric: Abstention rate (%) over all 6,069 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. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 2.5 Flash39#463
2Claude Sonnet 4.626.4#173
3Llama 4 Maverick18.8#1110
4Gemini Robotics-ER 1.616.5#170
5GPT-5.4 MiniLow14.8#406
6GPT-5.412.9#135
7GPT-5.4 Nano12.5#651
8Qwen 3.5 27B9.2#375

Interactive version: theaggregate.ai/benchmark?slug=roboabstention · How It Works · Data refreshed daily, snapshot 2026-11-21.