Beyond behavior: stimulus-response safety evaluation of vision-language-action models

Researchers conducted a stimulus-response safety evaluation of vision-language-action (VLA) models across three models and two video benchmarks, with results suggesting that even when danger is registered internally in the model, task execution may still be the default response. The evaluation examines the models' internal representations and action tendencies when facing visible danger and harmful consequences, and uses human biological sensitivity, in which safety-critical stimuli can trigger defensive responses, as a reference to test whether VLA models exhibit similar responses.