Research map / Video / Video-language understanding
LensWalk: Agentic Video Understanding by Planning How You See in Videos
Keliang Li, Yansong Li, Hongze Shen, Mengdi Liu, Hong Chang, Shiguang Shan
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 Scan Search tool (ablation) | LensWalk (o3/GPT-4.1) · Video-MME (long) | accuracy: -4.6% when removed | Improved |
| add Stitch Verify tool (ablation) | LensWalk (o3/GPT-4.1) · Video-MME (long) | accuracy: -3.2% when removed | Improved |
| add Segment Focus tool (ablation) | LensWalk (o3/GPT-4.1) · Video-MME (long) | accuracy: -1.9% when removed | Improved |
| add Timestamp Anchor (ablation) | LensWalk (o3/GPT-4.1) · Video-MME (long) | accuracy: to 69.7 | Improved |
| add Subject Memory Table (ablation) | LensWalk (o3/GPT-4.1) · Video-MME (long) | accuracy: 69.7 -> 70.0 | Improved |
| wrap o3 reasoner around GPT-4.1 observer | GPT-4.1 · Video-MME (long) | accuracy: 63.1 -> 70.0 (+6.9) | Improved |
| use open-source Qwen3-235B-A22B as reasoner instead of o3 (ablation) | LensWalk with GPT-4.1 observer · Video-MME (long) | accuracy: 63.1 -> 63.2 (+0.1) | No clear effect |
| use open-source Qwen3-235B-A22B as reasoner (ablation) | LensWalk with Qwen2.5-VL-72B observer · Video-MME (long) | accuracy: 63.1 -> 62.5 (-0.6) | Hurt |
| use open-source Qwen3-235B-A22B as reasoner (ablation) | LensWalk with Qwen2.5-VL-7B observer · Video-MME (long) | accuracy: 55.4 -> 59.7 (+4.3) | Improved |
| apply LensWalk with o3 as both reasoner and observer | o3 · LVBench | accuracy: 57.1 -> 68.6 | Improved |
| apply LensWalk with o3 as both reasoner and observer | o3 · Video-MME (long) | accuracy: 64.7 -> 71.4 | Improved |
| apply LensWalk with o3 reasoner and GPT-4.1 observer | o3 · MMVU (MC) | accuracy: 78.9 -> 80.9 | Improved |
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.
- Deep Video Discovery — Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding (NeurIPS 2025) · better on 2 of 3 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
- Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks (ICLR 2026 · similarity 0.89)
- Act2See: Emergent Active Visual Perception for Video Reasoning (CVPR 2026 · similarity 0.89)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.88)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.88)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.88)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (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.
