Research map / Video / Video-language understanding
WeaveTime: Streaming from Earlier Frames into Emergent Memory in VideoLLMs
Yulin Zhang, Cheng Shi, Sibei Yang
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 |
|---|---|---|---|
| finetune on offline video with timestamp prompts only (no temporal reconstruction) (ablation) | LLaVA-OV-7B + ReKV · OVO-Bench | overall accuracy (%): 53.56 -> 49.88 (-3.68) | Hurt |
| add Temporal Reconstruction auxiliary task (SOPE) to finetuning (ablation) | LLaVA-OV-7B + ReKV · OVO-Bench | overall accuracy (%): 49.88 -> 55.70 (+5.82) | Improved |
| add Temporal Reconstruction auxiliary task (SOPE) (ablation) | LLaVA-OV-7B + ReKV · Streaming-Bench Real-Time | accuracy (%): 65.91 -> 68.49 (+2.58) | Improved |
| add PCDF-Cache (uncertainty-gated recall) on top of SOPE (ablation) | LLaVA-OV-7B + ReKV · OVO-Bench | overall accuracy (%): 55.70 -> 57.57 (+1.87) | Improved |
| add PCDF-Cache on top of SOPE (ablation) | LLaVA-OV-7B + ReKV · Streaming-Bench Real-Time | accuracy (%): 68.49 -> 72.13 (+3.64) | Improved |
| use coarse-to-fine recall instead of coarse-only retrieval (ReKV) (ablation) | LLaVA-OV · QAEgo4D | recall: 23.9 -> 25.2 | Improved |
| use coarse-to-fine recall instead of coarse-only retrieval (ablation) | LLaVA-OV · MLVU | accuracy: 68.5 -> 68.9 | Improved |
| use fine-grained retrieval over full history instead of coarse-to-fine (ablation) | LLaVA-OV · QAEgo4D/MLVU/EventHALL | memory: OOM | Hurt |
| raise entropy threshold for recall in PCDF-Cache (ablation) | WeaveTime · OVO-Bench | overall accuracy / latency: accuracy peaks 57.57 at threshold 0.6; latency decreases monotonically | Mixed |
| add WeaveTime to LLaVA-OV-7B instead of ReKV | LLaVA-OV-7B · OVO-Bench Real-Time | average accuracy (%): 61.72 -> 68.82 | Improved |
| use SOPE with 30K offline samples instead of stream-tailored data | LLaVA-OV + ReKV · OVO-Bench | overall accuracy (%): 53.6 -> 55.7 with 8 GPUs and 0 stream-tailored data | Improved |
Closest papers by meaning
- What Should a Streaming Video Model Remember? (NeurIPS 2026 · similarity 0.91)
- Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs (NeurIPS 2026 · similarity 0.90)
- 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 0.89)
- CoRDS: Coreset-Based Representative and Diverse Selection for Streaming Video Understanding (NeurIPS 2026 · similarity 0.89)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 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.
