Research map / Video / Video-language understanding
Think While Watching: Online Streaming Segment-Level Memory for Multi-Turn Video Reasoning in Multimodal Large Language Models
Lu Wang, Zhuoran Jin, Yupu Hao, Yubo Chen, Kang Liu, Yulong Ao, Jun Zhao
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 segment memory notes (ablation) | TWW multi-turn S3 · StreamingBench (Qwen3-VL-4B) | accuracy: 52.35 -> 57.40 | Improved |
| use longer segments 120s/60s instead of 60s/30s (ablation) | TWW multi-turn S3 · StreamingBench | accuracy: 57.40 -> 55.33; tokens 302.56 -> 230.46 | Mixed |
| use shorter segments 30s/15s instead of 60s/30s (ablation) | TWW multi-turn S3 · StreamingBench | accuracy: 57.40 -> 57.20; tokens 302.56 -> 380.50 | No clear effect |
| add stage-3 long-range training over stage 2 (ablation) | TWW single-turn · OVO-Bench (Qwen3-VL-4B) | accuracy: 54.51 -> 55.02 | Improved |
| add stage-3 long-range training over stage 2 (ablation) | TWW single-turn · StreamingBench (Qwen3-VL-4B) | accuracy: 59.71 -> 60.04 | Improved |
| add streaming-aligned training to online Thinking (ablation) | Qwen3-VL-4B Thinking · StreamingBench | accuracy: 18.13 -> 60.04 | Improved |
| add streaming-aligned training (single-turn S3) | Qwen3-VL-4B Thinking · OVO-Bench | accuracy: 50.70 -> 55.02 | Improved |
| use multi-turn segment memory instead of Thinking | Qwen3-VL-4B · StreamingBench | avg output tokens: 689.22 -> 302.56 | Improved |
| use multi-turn segment memory instead of Thinking | Qwen3-VL-4B · StreamingBench | accuracy: 58.52 -> 57.40 | Hurt |
| use TWW pipeline instead of interleaved (ablation) | Qwen3-VL-4B · StreamingBench | accuracy: 55.35 -> 57.40 | Improved |
| use decoupled streaming pipeline instead of batch Thinking | Qwen3-VL-4B · StreamingBench | TTFT: 31203.69 -> 2304.28 | Improved |
| add single-turn streaming S3 training for offline Video-MME | Qwen3-VL-4B Thinking · Video-MME | accuracy: 68.89 -> 73.41 | Improved |
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.
- StreamForest-7B — StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025) · 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
- Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models (CVPR 2026 · similarity 0.93)
- Streaming Video Understanding and Multi-round Interaction with Memory-enhanced Knowledge (ICLR 2025 · similarity 0.91)
- StreamOV: Streaming Omni-Video Understanding via Evidence-Guided Memory and Response Triggering (NeurIPS 2026 · similarity 0.90)
- SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding (ICLR 2025 · similarity 0.90)
- StreamingThinker: Large Language Models Can Think While Reading (ICLR 2026 · similarity 0.90)
- What Should a Streaming Video Model Remember? (NeurIPS 2026 · 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.
