Research map / Video / Video-language understanding
TimeLens: Rethinking Video Temporal Grounding with Multimodal LLMs
Jun Zhang, Teng Wang, Yuying Ge, Yixiao Ge, Xinhao Li, Limin Wang
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 |
|---|---|---|---|
| re-annotate benchmarks with strict criteria | VTG models · Charades-STA | Mean IoU: re-ranking | Mixed |
| use interleaved textual timestamp prefix with raw timestamps instead of visual overlay (ablation) | Qwen2.5-VL-7B · ActivityNet-TimeLens | mIoU: 39.8 -> 43.1 | Improved |
| use interleaved rather than non-interleaved textual prefix (ablation) | Qwen2.5-VL-7B · QVHighlights-TimeLens | mIoU: 42.8 -> 56.7 | Improved |
| use raw timestamps instead of frame indices (ablation) | Qwen2.5-VL-7B · ActivityNet-TimeLens | mIoU: 36.9 -> 43.1 | Improved |
| replace position-embedding time encoding with interleaved text (ablation) | Qwen2.5-VL-7B · Charades-TimeLens | mIoU: 36.6 -> 48.3 | Improved |
| thinking-free RLVR instead of SFT on 100K data (ablation) | Qwen2.5-VL-7B · ActivityNet-TimeLens | mIoU: 39.7 -> 43.1 | Improved |
| remove thinking from RLVR (ablation) | Qwen2.5-VL-7B · Charades-TimeLens | mIoU: 42.7 -> 48.3 | Improved |
| add SFT phase before RLVR (ablation) | Qwen2.5-VL-7B · Charades-TimeLens | mIoU: 48.3 -> 50.1 | No clear effect |
| continue RLVR training past reward plateau (ablation) | Qwen2.5-VL-7B · TimeLens-Bench | performance | Hurt |
| sample higher-difficulty training data for RLVR (ablation) | Qwen2.5-VL-7B · TimeLens-Bench | performance | Improved |
| train on TimeLens-100K re-annotated data | Qwen2.5-VL-7B · TimeLens-Bench | performance | Improved |
Later papers that compared with this method
2 later accepted papers report TimeLens-8B as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation (ICML 2026 · better on 0 of 1 shared result)
- Towards One-to-Many Temporal Grounding (ICML 2026 · better on 1 of 1 shared result)
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
- VTG-LLM: Integrating Timestamp Knowledge into Video LLMs for Enhanced Video Temporal Grounding (AAAI 2025 · similarity 0.92)
- Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.91)
- TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning (ICLR 2025 · similarity 0.91)
- ReVisionLLM: Recursive Vision-Language Model for Temporal Grounding in Hour-Long Videos (CVPR 2025 · similarity 0.90)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.90)
- TRACE: Temporal Grounding Video LLM via Causal Event Modeling (ICLR 2025 · similarity 0.90)
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.
