Research map / Video / Video-language understanding
StreamingVLM: Real-Time Understanding for Infinite Video Streams
Ruyi Xu, Guangxuan Xiao, Yukang Chen, Liuning He, Kelly Peng, Yao Lu, Song Han
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 overlapped-chunk full-attention SFT on Inf-Streams-Train (vs Live-WhisperX-526K only) (ablation) | Qwen2.5-VL-7B · Inf-Streams-Eval win rate vs GPT-4o mini | win rate: 32.17 -> 63.46 | Improved |
| add high-quality annealing data (ablation) | StreamingVLM · Inf-Streams-Eval win rate vs GPT-4o mini | win rate: 63.46 -> 66.18 | Improved |
| use overlapped chunks instead of non-overlapping chunks in SFT (ablation) | StreamingVLM · Inf-Streams-Eval win rate vs GPT-4o mini | win rate: 62.51 -> 66.18 | Improved |
| use overlapped chunks instead of non-overlapping chunks in SFT (ablation) | StreamingVLM · VideoMME (w/o sub.) | accuracy: 65.50 -> 65.10 | No clear effect |
| use overlapped chunks instead of non-overlapping chunks in SFT (ablation) | StreamingVLM · OVOBench (Realtime) | accuracy: 59.20 -> 61.96 | Improved |
| evict old text tokens (sink + window) instead of keeping all text (ablation) | StreamingVLM (basketball subset) · Inf-Streams-Eval Basketball win rate vs GPT-4o mini | win rate: 60.41 -> 73.64 | Improved |
| keep a recent vision window (16 s vs 0 s) (ablation) | StreamingVLM · Inf-Streams-Eval win rate vs GPT-4o mini | win rate: 52.90 -> 66.18 | Improved |
| enlarge vision window from 16 s to 32 s (ablation) | StreamingVLM · Inf-Streams-Eval win rate vs GPT-4o mini | win rate: 66.18 -> 65.49 | No clear effect |
| keep text attention sink (T_sink=512 vs 0) (ablation) | StreamingVLM (basketball subset) · Inf-Streams-Eval Basketball win rate vs GPT-4o mini | win rate: 66.76 -> 73.64 | Improved |
| increase T_sink from 64 to 1024 (ablation) | StreamingVLM · Inf-Streams-Eval win rate vs GPT-4o mini | win rate: 72.04 -> 74.82 | Improved |
| use ReKV eviction with the fine-tuned StreamingVLM instead of proposed KV cache (ablation) | StreamingVLM · Inf-Streams-Eval win rate | win rate: 66.18 -> 0.00 | Hurt |
| SFT on Live-WhisperX-526K only versus base model on VQA (ablation) | Qwen2.5-VL-7B · MVBench | accuracy: 67.34 -> 63.71 | Hurt |
Later papers that compared with this method
2 later accepted papers report StreamingVLM as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Don't Pause! Every prediction matters in a streaming video (NeurIPS 2026 · better on 2 of 2 shared results)
- Streaming Video Instruction Tuning (CVPR 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
- Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models (CVPR 2026 · similarity 0.91)
- StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025 · similarity 0.90)
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.90)
- SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding (ICLR 2025 · similarity 0.89)
- VideoLLM-MoD: Efficient Video-Language Streaming with Mixture-of-Depths Vision Computation (NeurIPS 2024 · similarity 0.89)
- CoRDS: Coreset-Based Representative and Diverse Selection for Streaming Video Understanding (NeurIPS 2026 · 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.
