Megadose The AI curve, compressed.

AmbiSteer: Selectively Eliciting Ambiguity-Aware Responses in Vision-Language Models

· ArXiv · AI/CL/LG ·
AmbiSteer raised appropriate handling of visual questions from 59% to 77% on average across three models.

The study examines questions where several entities in an image could plausibly be the subject. Prompting generally worked better than steering model activations alone at eliciting clarification or lists of possible referents. Combining the two increased those responses, but also triggered them on unambiguous questions. AmbiSteer uses a linear classifier to predict ambiguity and condition the steering direction, helping models distinguish when to clarify and when to answer directly.

ArXiv · AI/CL/LG's note

score 5

Categories: Research