Self‑grounded verification reduces agreement bias in multimodal LLMs

Moises Andrade and colleagues present SGV, a verification technique for multimodal LLMs that first asks the model to list what correct behavior looks like, then judges a specific action against that list. The paper shows the method cuts the models’ tendency to over‑approve actions – called agreement bias – and raises failure‑detection rates by up to 25 percentage points and overall accuracy by 14 percentage points. Applied to a GUI specialist in OSWorld, a diffusion policy in RoboMimic, and a ReAct agent in VisualWebArena, the verifiers improve task completion by about 20 percentage points over the previous best results.