AI Paper Atlas

Research map / Video / Video-language understanding

Progressive Online Video Understanding with Evidence-Aligned Timing and Transparent Decisions

Kecheng Zhang, Zongxin Yang, Mingfei Han, Haihong Hao, Yunzhi Zhuge, Changlin Li, junhan zhao, Zhihui Li, Xiaojun Chang

ICLR 2026 · Video › Video-language understanding · approach: Event and anomaly reasoning in video · 1 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
add HPSI multi-level aggregation instead of single-shot adaptive pooling before the LLM (ablation)Thinking-QwenVL · OVOBenchoverall accuracy: 43.4 -> 46.9 (-3.5 when pooled)Improved
add HPSI multi-level aggregation instead of single-shot adaptive pooling before the LLM (ablation)Thinking-QwenVL · VideoMMEoverall accuracy: 48.9 -> 56.3 (-7.4 when pooled)Improved
add the first-frame token (ablation)Thinking-QwenVL · OVOBenchoverall accuracy: 42.6 -> 46.9Improved
add level-3 aggregation tokens (ablation)Thinking-QwenVL · OVOBenchoverall accuracy: 46.0 -> 46.9 on OVOBench; 56.0 -> 56.3 on VideoMMEMixed
add level-2 and level-3 aggregation tokens (ablation)Thinking-QwenVL · OVOBenchoverall accuracy: 49.6 -> 46.9 on OVOBench; 54.7 -> 56.3 on VideoMMEMixed
add ATDM decision making to Flash-VStreamFlash-VStream · StreamingBenchaccuracy: 22.53 -> 26.58Improved
replace graded progress/confidence (rho, c) with a binary answerable flag (ablation)Thinking-QwenVL · online benchmarksaccuracy: 3.62% accuracy drop when removedImproved
add question decomposition (P2) (ablation)Flash-VStream · online benchmarksaccuracy: 3.60% accuracy drop when removedImproved
add streaming captioning (P3) (ablation)Flash-VStream · online benchmarksaccuracy: 4.49% accuracy drop when removedImproved
use 3D RoPE instead of offset sequential positional embedding (OSPR) (ablation)Thinking-QwenVL · OVOBenchoverall accuracy: 43.3 -> 46.9Improved
use full-frame (100%) attention redistribution rather than 93.75% frame reductionThinking-QwenVL · OVOBenchoverall accuracy: 46.9 -> 52.5Improved
apply HPSI and ATDM to Qwen2.5-VL baselineQwen2.5-VL-7B · RTVBenchoverall accuracy: 32.75 -> 35.87Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy