Research map / Video / Video-language understanding
Temporal-Aware Reasoning Optimization for Video Temporal Grounding
Minghang Zheng, Zihao Yin, Yi Yang, Yuxin Peng, Yang 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 |
|---|---|---|---|
| use TaRO instead of Time-R1 (Qwen2.5-VL-7B) | Time-R1 · Charades-STA, TVGBench zero-shot | R1@0.5: 60.8 -> 64.8 (Charades-STA), 29.4 -> 37.8 (TVGBench) | Improved |
| add temporal-sensitive reward alone to the RL baseline (ablation) | GRPO baseline · Charades-STA zero-shot | R1@0.5: 61.1 -> 63.1 | Improved |
| use constructive reasoning exploration alone without the self-exploration phase (ablation) | TaRO · Charades-STA zero-shot | R1@0.5: 61.1 -> 51.6 | Hurt |
| add constructive reasoning exploration with progressive curriculum (ablation) | TaRO without temporal reward · Charades-STA zero-shot | R1@0.5: 61.1 -> 63.9 | Improved |
| add all three components (CRE, TR, PC) (ablation) | standard RL baseline · Charades-STA zero-shot | R1@0.5: 61.1 -> 64.8 | Improved |
| use constructed reasoning with RL-style AW-BC instead of SFT (ablation) | CRE warm-up · Charades-STA zero-shot | R1@0.5: 63.3 -> 64.8 | Improved |
| sample a subset of captions instead of using all captions (ablation) | CRE · Charades-STA zero-shot | R1@0.5: 63.6 -> 64.8 | Improved |
| use AW-BC loss instead of GRPO loss on constructed off-policy rollouts (ablation) | CRE warm-up · Charades-STA zero-shot | R1@0.5: 46.7 -> 64.8 | Improved |
| add advantage weighting and positive-advantage filter to behavioral cloning (ablation) | AW-BC loss · Charades-STA zero-shot | R1@0.7: 33.7 -> 38.4 (filter), 37.1 -> 38.4 (weighting) | Improved |
| use a weaker open-source captioner (Qwen3.5-4B or 9B) instead of Gemini-3-Pro (ablation) | CRE · Charades-STA zero-shot | R1@0.5: 64.8 -> 64.2 (9B), 64.3 (4B) | No clear effect |
| shuffle only frames near ground-truth boundaries instead of the full video or random cropping (ablation) | temporal-sensitive reward · Charades-STA zero-shot | R1@0.5: 63.5 -> 64.8 (full shuffle), 63.1 -> 64.8 (cropping) | Improved |
| lower IoU gate threshold tau to 0.3 or remove gate effect with tau=1.0 (ablation) | temporal-sensitive reward · Charades-STA zero-shot | R1@0.5: 64.8 -> 61.8 (tau=0.3), 63.9 (tau=1.0) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Time-R1 — Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025) · better on 8 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TAR: Temporal Anchor-Constrained Reasoning for Video Temporal Grounding (ECCV 2026 · similarity 0.93)
- VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting (ECCV 2026 · similarity 0.89)
- ReXTime: A Benchmark Suite for Reasoning-Across-Time in Videos (NeurIPS 2024 · similarity 0.88)
- Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.88)
- Momentor: Advancing Video Large Language Model with Fine-Grained Temporal Reasoning (ICML 2024 · similarity 0.88)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (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.
