Clarify or Answer: Reinforcement Learning for Agentic VQA with Context Under-specification
Aug 1, 2026·
·
1 min read
Zongwan Cao
Equal contribution
Bingbing Wen
Equal contribution
,Lucy Lu Wang
Abstract
We propose a reinforcement learning framework for agentic visual question answering (VQA) under context under-specification, enabling agents to decide when to clarify missing information versus directly answer.
Type
Publication
COLM 2026
We introduce a reinforcement learning framework for agentic VQA that explicitly models whether an agent should ask for clarification or answer directly when faced with underspecified visual and textual context.
