Research map / Video / Video-language understanding
When Thinking Hurts: Mitigating Visual Forgetting in Video Reasoning via Frame Repetition
Xiaokun Sun, Yubo Wang, Haoyu Cao, Linli Xu
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 |
|---|---|---|---|
| add FrameRepeat (repeat 8 learned key frames) | Qwen2.5-VL-7B (128 frames) · VideoMME | accuracy: 60.9 -> 63.4 | Improved |
| add FrameRepeat (repeat 8 learned key frames) | Qwen2.5-VL-7B (128 frames) · LongVideoBench | accuracy: 51.4 -> 55.1 | Improved |
| add FrameRepeat to Qwen3-VL-8B-Thinking | Qwen3-VL-8B-Thinking (128 frames) · LVBench | accuracy: 40.6 -> 44.3 | Improved |
| add FrameRepeat to Qwen3-VL-8B-Thinking, scorer trained only on Qwen2.5-VL | Qwen3-VL-8B-Thinking (128 frames) · MLVU | accuracy: 68.2 -> 71.2 | Improved |
| repeat Bottom-K frames instead of base | Qwen3-VL-8B-Thinking (128 frames) · VideoMME | accuracy: 67.0 -> 66.3 | Hurt |
| repeat Random-K frames | Qwen2.5-VL-7B (128 frames) · MLVU | accuracy: 63.0 -> 61.5 | Hurt |
| add ranking loss to regression loss (ablation) | FrameRepeat Qwen2.5-VL · VideoMME | accuracy: 62.1 -> 63.4 | Improved |
| add ranking loss to regression loss (ablation) | FrameRepeat Qwen3-VL · Video-Holmes | accuracy: 42.1 -> 43.4 | Improved |
| increase repeated frames K from 4 to 8 (ablation) | FrameRepeat Qwen2.5-VL · VideoMME | accuracy: 62.3 -> 63.4 | Improved |
| increase repeated frames K from 8 to 16 (ablation) | FrameRepeat Qwen2.5-VL · VideoMME | accuracy: 63.4 -> 62.9 | Hurt |
| use learned FrameRepeat scores instead of TSPO scores for repetition (ablation) | Qwen2.5-VL-7B · VideoMME | accuracy: 62.8 -> 63.4 | Improved |
| repeat Top-4 frames in thinking mode instead of no repetition (ablation) | Qwen2.5-VL (128 frames) · VideoMME subset (300 questions) | accuracy: +2.3% (Top-4 65.0%) | Improved |
Closest papers by meaning
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.90)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.89)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.89)
- Not Another Text Benchmark: Putting the “Visual" Back in Visual Question Answering for Large Video Models (NeurIPS 2026 · similarity 0.89)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.88)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · 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.
