Research map / Video / Video-language understanding
Thinking in Streaming Video
Zikang Liu, Longteng Guo, Handong Li, Ru Zhen, Xingjian He, Ruyi Ji, Xiaoming Ren, Yanhao Zhang, Haonan Lu, Jing 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 |
|---|---|---|---|
| increase reasoning token budget from 0 to 20 tokens per second (ablation) | ThinkStream-3B · OVO-Bench Backward | accuracy: 41.8 -> 52.3 | Improved |
| raise reasoning budget from 20 to 30 tokens per second (ablation) | ThinkStream-3B · OVO-Bench Backward | accuracy: 52.3 -> 52.6, latency 380 ms -> 505 ms | No clear effect |
| use 20s instead of 5s visual KV window (ablation) | ThinkStream-3B · OVO-Bench Backward | accuracy: 46.7 -> 52.3 | Improved |
| expand visual window from 20s to 30s (ablation) | ThinkStream-3B · OVO-Bench Backward | accuracy: 52.3 -> 51.6 | Hurt |
| add cold-start CoT memory versus no memory (ablation) | Qwen2.5-VL-3B streaming · Memory variants average | average score: 56.9 -> 60.5 | Improved |
| use discrete caption tokens as memory instead of no memory (ablation) | Qwen2.5-VL-3B streaming · Memory variants average | average score: 56.9 -> 48.7 | Hurt |
| add RLVR post-training on top of cold-start CoT memory (ablation) | ThinkStream-3B · Memory variants average | average score: 60.5 -> 64.8 | Improved |
| apply ThinkStream framework to Qwen2.5-VL-3B | Qwen2.5-VL-3B · OVO-Bench | overall average accuracy: 51.00 -> 59.66 | Improved |
| apply ThinkStream framework to Qwen2.5-VL-7B | Qwen2.5-VL-7B · OVO-Bench Backward Tracing | average accuracy: 52.28 -> 60.68 | Improved |
| apply ThinkStream with visual token eviction to offline benchmarks | Qwen2.5-VL-3B · VideoMME and LongVideoBench | accuracy: 61.5 -> 61.9 (VideoMME), 54.2 -> 56.4 (LongVideoBench) | Improved |
| CUDA Graph streaming inference engine versus standard Qwen2.5-VL-3B decoding at batch size 1 | ThinkStream-3B · Token decoding speed | tokens/s: 30.06 -> 154.07 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- Dispider-7B — Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction (CVPR 2025) · better on 2 of 2 shared results
- StreamForest-7B — StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025) · better on 1 of 1 shared result
- Flash-VStream-7B — Flash-VStream: Efficient Real-Time Understanding for Long Video Streams (ICCV 2025) · better on 1 of 1 shared result
- VideoLLM-online-8B — VideoLLM-online: Online Video Large Language Model for Streaming Video (CVPR 2024) · 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 0.91)
- Streaming Video Understanding and Multi-round Interaction with Memory-enhanced Knowledge (ICLR 2025 · similarity 0.90)
- What Should a Streaming Video Model Remember? (NeurIPS 2026 · similarity 0.90)
- StreamingThinker: Large Language Models Can Think While Reading (ICLR 2026 · similarity 0.89)
- StreamOV: Streaming Omni-Video Understanding via Evidence-Guided Memory and Response Triggering (NeurIPS 2026 · 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.
