VideoSearch-R1 presents an agentic framework that jointly handles inter-video retrieval and intra-video reasoning, including query-conditioned temporal grounding. Its Soft Query Refinement (SQR) method adjusts query tokens in a continuous latent space instead of repeatedly rewriting discrete text queries. The retrieval-and-reasoning process is trained with Group Relative Policy Optimization (GRPO), using task-level rewards from both retrieval and downstream objectives. The paper reports state-of-the-art results on Video Corpus Moment Retrieval (VCMR) across three datasets, and states that code and model checkpoints are publicly available.
No heat snapshots are available in the last 24 hours.