RoboAbstention benchmark: VLMs fail to say 'no' in robot tasks

A new paper by researchers at Purdue University, titled 'The Yes-Man Syndrome,' highlights a critical flaw in vision-language models (VLMs) used as robotic planners: they rarely refuse instructions that are ambiguous, physically impossible, or based on false premises. The team created RoboAbstention, a scalable framework that generates instructions grounded in images from five robotics datasets. The pipeline includes structured visual grounding, deterministic constraint derivation, and controlled instruction generation via category-specific templates, producing a diverse dataset of 6,069 instructions with verifiable abstention conditions.

Testing frontier VLMs revealed severe weaknesses. Gemini 2.5 Flash, the best-performing model, abstained on only 39.0% of instructions. The embodied robotics planner Gemini Robotics ER 1.6 Preview abstained on just 16.5%. Even models with advanced reasoning capabilities struggled. However, interventions like defensive prompting and in-context learning dramatically improved performance: Gemini Robotics ER 1.6 Preview reached 93.6% abstention, and GPT-5.4 Mini reached 88.6%. The open-source dataset and framework are available on GitHub, aiming to help build safer, more trustworthy embodied agents.