AI Paper Atlas

Research map / Video / Video-language understanding

TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning

Junwen Pan, Qizhe Zhang, Rui Zhang, Ming Lu, Xin Wan, Yuan Zhang, Chang Liu, Qi She

ICLR 2026 · Video › Video-language understanding · approach: Long video understanding · 16 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add RL post-training with GRPO-CSV after SFT (ablation)TimeSearch-R · VideoMME (w/o sub)overall accuracy: 59.2 -> 66.6Improved
Add SFT cold start to zero-shot CoT search (ablation)TimeSearch-R · Temporal searchtemporal F1: 0.0 -> 7.8Improved
Add SFT cold start to zero-shot CoT search (ablation)TimeSearch-R · Temporal searchsearched frame completeness: 44.2% -> 60.5%Improved
Add RL after SFT, effect on reasoning consistency (ablation)TimeSearch-R · Reasoning consistencyconsistency rate: +2.6%Improved
Add CSV to GRPO (completeness) (ablation)GRPO · Temporal searchcompleteness: 57.2% -> 60.5%Improved
Add CSV to GRPO (temporal F1) (ablation)GRPO · Temporal searchtemporal F1: 7.4 -> 7.8Improved
Add CSV to GRPO, training stability (ablation)GRPO · Training dynamicscompleteness: training collapses around step 300 without CSV; completeness drops to zeroImproved
Add accuracy reward to completeness-only reward (ablation)GRPO-CSV · QA performanceQA accuracy: completeness reward alone reduces QA accuracy by 0.3%Improved
Apply two-stage data filtering in RL training (Ego+Exo) (ablation)TimeSearch-R · VideoMME (w/o sub)overall accuracy: 62.8 -> 66.6Improved
Add exocentric data to filtered egocentric data (ablation)TimeSearch-R · VideoMME (w/o sub)overall accuracy: 65.3 -> 66.6Improved
Add search without training via CoT prompting to Qwen2.5-VLQwen2.5VL-7B · VideoMME (w/o sub)overall accuracy: 65.1 -> 51.8Hurt
Train with TimeSearch-R instead of base Qwen2.5-VLQwen2.5VL-7B · LongVideoBench valaccuracy: 56.0 -> 60.1Improved

Later papers that compared with this method

2 later accepted papers report TimeSearch-R-7B as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.

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.

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