# Towards Temporal Compositional Reasoning in Long-Form Sports Videos

Source: [Towards Temporal Compositional Reasoning in Long-Form Sports Videos](<https://www.sota2.com/research/paper/69eeed9555a3c32af89f97b1>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-07-15T05:00:05.882Z

Authors: Siyu Cao, Lu Zhang, Ruizhe Zeng, Zhi-yong Liu

Published: 2026-04-24T05:02:03.000Z

## Abstract

Sports videos are a challenging domain for multimodal understanding because they involve complex and dynamic human activities. Despite rapid progress in Multimodal Large Language Models (MLLMs), long-horizon reasoning in sports videos remains difficult, as answering questions requires both locating temporally sparse evidence and integrating it into reasoning. We attribute this limitation to two closely coupled factors: insufficient supervision over temporally dispersed evidence, and the lack of methods that require models to identify, localize, and justify temporal evidence. To address these gaps, we introduce SportsTime, a large-scale benchmark for long-form sports video understanding, comprising 14K+ open-ended QA pairs and 50K+ step-wise temporal evidence annotations. Building on SportsTime, we propose Chain-of-Time Reasoning (CoTR), which treats reasoning as a process of temporally grounded evidence composition. Specifically, during training, CoTR introduces a temporal-reward GRPO to encourage temporally grounded reasoning. During inference, it employs an anchor-observe-infer evidence-seeking loop to iteratively localize, verify, and compose temporal evidence before producing the final answer. Experiments demonstrate the usefulness of SportsTime as a benchmark and the effectiveness of CoTR, which consistently improves temporal compositional reasoning and step-wise grounding quality over strong MLLM baselines.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Sports Video Question Answering on SportsTime 200 stratified samples (test)](<https://www.sota2.com/research/sota/sports-video-question-answering-on-sportstime-200-stratified-samples-test.md>): Sports Video Question Answering; SportsTime 200 stratified samples (test); Human Average Score: 30.5.

- [Sports Video Question Answering on SportsTime](<https://www.sota2.com/research/sota/sports-video-question-answering-on-sportstime.md>): Sports Video Question Answering; SportsTime; Perception Score: 25.12.

- [Video Understanding on LVBench](<https://www.sota2.com/research/sota/video-understanding-on-lvbench.md>): Video Understanding; LVBench; Overall Accuracy: not recorded.

- [Video Understanding on MLVU (Score, Avg Score, Acc. %)](<https://www.sota2.com/research/sota/video-understanding-on-mlvu-score-avg-score-acc-percent.md>): Video Understanding; MLVU; Accuracy: not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.22226>)

[PDF](<https://arxiv.org/pdf/2604.22226.pdf>)