Research map / Video / Video-language understanding
Towards Temporal Compositional Reasoning in Long-Form Sports Videos
Siyu Cao, Lu Zhang, Ruizhe Zeng, Zhi-yong Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with direct SFT (ablation) | Qwen3-VL-4B · SportsTime (Avg) | open-ended QA accuracy (%): 25.15 -> 17.61 | Hurt |
| fine-tune with native GRPO (ablation) | Qwen3-VL-4B · SportsTime (Avg) | open-ended QA accuracy (%): 25.15 -> 26.72 | Improved |
| fine-tune with native GRPO (ablation) | Qwen3-VL-4B · SportsTime (Temporal) | open-ended QA accuracy (%): 13.14 -> 11.76 | Hurt |
| replace native GRPO with temporal-reward GRPO (tr-GRPO w/ ts) (ablation) | Qwen3-VL-4B · SportsTime (Avg) | open-ended QA accuracy (%): 26.72 -> 27.31 | Improved |
| replace native GRPO with temporal-reward GRPO (tr-GRPO w/ ts) (ablation) | Qwen3-VL-4B · SportsTime (Tactical) | open-ended QA accuracy (%): 32.65 -> 35.73 | Improved |
| add timestamp overlay to frames in tr-GRPO (ablation) | Qwen3-VL-4B · SportsTime (Avg) | open-ended QA accuracy (%): 26.12 -> 27.31 | Improved |
| add anchor-triggered interactive observation (AT-IO) at inference (ablation) | Qwen3-VL-4B · SportsTime (Avg) | open-ended QA accuracy (%): 27.31 -> 29.24 | Improved |
| add anchor-triggered interactive observation (AT-IO) at inference (ablation) | Qwen3-VL-4B · SportsTime (Temporal) | open-ended QA accuracy (%): 13.73 -> 15.43 | Improved |
| add anchor-triggered interactive observation (AT-IO) at inference (ablation) | Qwen3-VL-4B · SportsTime (Tactical) | open-ended QA accuracy (%): 35.73 -> 34.55 | Hurt |
| add tr-GRPO reward over native GRPO for temporal grounding | Qwen3-VL-4B · SportsTime (SGA objective) | mIoU / Hit@0.5: 0.12 -> 0.58 | Improved |
| prompt zero-shot CoT for time spans instead of no time prompt | Qwen3-VL-4B · SportsTime (SGA objective) | anchor rate (%): 19.49% anchor, mIoU 0.12 | Mixed |
| apply CoTR on general long-video benchmarks | Qwen3-VL-4B · LVBench | accuracy (%): 56.2 -> 59.9 | Improved |
Closest papers by meaning
- SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports (ICLR 2026 · similarity 0.91)
- SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models (ICLR 2025 · similarity 0.90)
- VELOCITI: Benchmarking Video-Language Compositional Reasoning with Strict Entailment (CVPR 2025 · similarity 0.89)
- Temporal Reasoning Transfer from Text to Video (ICLR 2025 · similarity 0.88)
- STEP: Enhancing Video-LLMs' Compositional Reasoning by Spatio-Temporal Graph-guided Self-Training (CVPR 2025 · similarity 0.88)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
