Researchers introduce SportsTime benchmark and Chain-of-Time Reasoning method for temporal understanding in long-form sports videos

SportsTime is a large-scale benchmark comprising 14K+ open-ended QA pairs and 50K+ step-wise temporal evidence annotations designed for long-form sports video understanding.

Chain-of-Time Reasoning (CoTR) treats reasoning as temporally grounded evidence composition, using a temporal-reward GRPO (a training technique) during training and an anchor-observe-infer loop during inference to iteratively localize, verify, and compose temporal evidence before producing answers.

Experiments show CoTR consistently improves temporal compositional reasoning and step-wise grounding quality over strong MLLM (Multimodal Large Language Model) baselines.