New research reveals that AI models overwhelmingly prefer to guess rather than request missing information, but reinforcement learning offers a potential solution.

ProactiveBench testing framework evaluated 22 multimodal language models on their ability to ask for help when visual information is unavailable

Nearly all 22 models tested failed to request needed information, instead choosing to make inferences despite incomplete data

Researchers demonstrated that simple reinforcement learning approaches can effectively train models to proactively ask users for clarification when needed