Towards One-to-Many Temporal Grounding

Towards One-to-Many Temporal Grounding, from a team including authors at Wuhan University, Peking University, Nanyang Technological University, and the National University of Singapore, formalizes a new video understanding problem where one text query must retrieve all disjoint matching moments in a video, rather than a single segment. The paper argues that existing MLLMs such as Gemini 2.5 Pro, Gemini 3 Pro, Seed-1.8, and Qwen3-VL fail because they lack event-cardinality perception, often merging multiple occurrences into one span or hallucinating extra segments. To measure this setting correctly, the authors introduce the OMTG Bench, a 340-sample benchmark with manually verified multi-segment annotations, plus new metrics: Count Accuracy, Temporal F1, and Effective Temporal F1, which zeroes out examples with wrong segment counts. They also build a 56k-sample instruction-tuning dataset using a five-stage pipeline based on Qwen3-VL-235B and Gemini 2.5 Pro, including repetitive-event discovery, grounding, strict visual veto filtering, recall checking, query refinement, and query-guided dense captioning. Training combines supervised fine-tuning with GRPO reinforcement learning using a composite reward that explicitly optimizes temporal IoU, count accuracy, caption quality, and output length. On OMTG Bench, their OMTG-4B model reaches 43.65 EtF1, outperforming Gemini 2.5 Pro by 15.85 points and Seed-1.8 by 15.61 points, while also improving standard one-to-one grounding on TimeLens and general video understanding on VideoMME.