Research map / Video / Video-language understanding
VideoTIR: Accurate Understanding for Long Videos with Efficient Tool-Integrated Reasoning
Zhe Gao, Shiyu Shen, Taifeng Chai, Weinong Wang, Haotian Xu, Xing W, Wenbin Li, Qi Fan, Yang Gao, Dacheng Tao
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 |
|---|---|---|---|
| replace GRPO with TAGPO (ablation) | VideoTIR 7B · Validation at 30 training steps | val accuracy: 21.1 -> 24.6 | Improved |
| replace GRPO with TAGPO (ablation) | VideoTIR 7B (8-frame Low) · Video-MME w/o sub, MVBench, LVB val | accuracy: Video-MME 54.1 -> 54.2; MVBench 56.9 -> 56.0; LVB 50.9 -> 51.3 | Mixed |
| add RL (GRPO) on top of zero-shot tool use (ablation) | Qwen2.5-VL-7B · Video-MME w/o sub (8-frame Low) | accuracy: 42.9 -> 54.1 | Improved |
| add tool calling (zero-shot) to thinking Qwen2.5-VL-7B at low res (ablation) | Qwen2.5-VL-7B · Video-MME w/o sub (8-frame Low) | accuracy: 35.0 -> 42.9 | Improved |
| increase input to 16 frames at high resolution (ablation) | VideoTIR (TAGPO) · Video-MME w/o sub (Mean) | accuracy: 54.2 -> 57.9 | Improved |
| enable thinking mode on Qwen2.5-VL-7B (ablation) | Qwen2.5-VL-7B · Video-MME w/o sub (Mean) | accuracy: 53.8 -> 49.5 | Hurt |
| use Random-Noised SFT (vision masked) instead of normal SFT for cold start (ablation) | Qwen2.5-VL-3B · MVBench (10% subset) | accuracy: 36.3 -> 40.7 | Improved |
| use Random-Noised SFT instead of normal SFT (ablation) | Qwen2.5-VL-3B · Video-MME (10% subset) | accuracy: 35.0 -> 38.5 | Improved |
| add SFT cold start with synthetic trajectories (ablation) | Qwen2.5-VL-3B · Format quality | format quality: <=1.0 -> 90.1 (Long prompt) | Improved |
| use internal toolkit instead of external tool pipelines | VideoTIR · Efficiency | latency (s): 68.3s (COLT) -> 4.6s | Improved |
| VideoTIR 16-frame High vs base Qwen2.5-VL-7B | Qwen2.5-VL-7B · MVBench | accuracy: 60.6 -> 64.3 | Improved |
| Zero-RL directly on 3B model | Qwen2.5-VL-3B · Zero-RL training | valid tool calls | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Video-MTR — Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding (ICML 2026) · better on 1 of 2 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
- VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding (CVPR 2025 · similarity 0.91)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.91)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.90)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.89)
- Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering (NeurIPS 2025 · similarity 0.89)
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.
