Research map / Video / Video-language understanding
Native Active Perception as Reasoning for Omni-Modal Understanding
Zhenghao Xing, Ruiyang Xu, Yuxuan Wang, Jinzheng He, Ziyang Ma, Qize Yang, Yunfei Chu, Jin Xu, Junyang Lin, Chi-Wing Fu, Pheng-Ann Heng
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 |
|---|---|---|---|
| fine-tune on static QA pairs (standard SFT) instead of agentic trajectories (ablation) | Qwen2.5-Omni-7B · LVBench | accuracy: 43.0 -> 41.6 | Hurt |
| use Agentic SFT (OTA trajectories) instead of standard SFT (ablation) | Qwen2.5-Omni-7B · LVBench | accuracy: 41.6 -> 48.7 | Improved |
| add Agentic SFT to base model (ablation) | Qwen2.5-Omni-7B · DailyOmni | accuracy: 60.1 -> 63.3 | Improved |
| add vanilla GRPO RL after Agentic SFT (ablation) | OmniAgent · DailyOmni | accuracy: 63.3 -> 62.2 | Hurt |
| add vanilla GRPO RL after Agentic SFT (ablation) | OmniAgent · MLVU | accuracy: 69.9 -> 69.9 | No clear effect |
| add vanilla GRPO RL after Agentic SFT (ablation) | OmniAgent · LVBench | accuracy: 48.7 -> 49.8 | Improved |
| use TAURA turn-level entropy rescaled advantage instead of vanilla GRPO (ablation) | OmniAgent · MLVU | accuracy: 69.9 -> 71.1 | Improved |
| use TAURA instead of vanilla GRPO (ablation) | OmniAgent · DailyOmni | accuracy: 62.2 -> 64.8 | Improved |
| use TAURA instead of vanilla GRPO (ablation) | OmniAgent · LVBench | accuracy: 49.8 -> 50.5 | Improved |
| increase maximum turn limit K from 6 to 52 (test-time scaling) (ablation) | OmniAgent · VideoMME-Long | accuracy: 53.4 -> 59.6 (+6.2%) | Improved |
| use active on-demand perception instead of dense passive sampling (vs Qwen2.5-VL-72B) | OmniAgent-7B · LVBench | accuracy: 47.3 -> 50.5 with 203 vs 768 frames | Improved |
| longer videos (about 30 to 130 min) with active perception (ablation) | OmniAgent · LVBench by duration | accuracy: 53.7 (20-40 min) -> 50.8 (120-140 min) | No clear effect |
Closest papers by meaning
- A Benchmark for Omni-Modal Reasoning in Long Videos (NeurIPS 2026 · similarity 0.91)
- Omni-R1: Reinforcement Learning for Omnimodal Reasoning via Two-System Collaboration (NeurIPS 2025 · similarity 0.89)
- Act2See: Emergent Active Visual Perception for Video Reasoning (CVPR 2026 · similarity 0.89)
- Omni-Perception Policy Optimization for Multimodal Emotion Reasoning (ICML 2026 · similarity 0.87)
- V-ABS: Action-Observer Driven Beam Search for Dynamic Visual Reasoning (ICML 2026 · similarity 0.87)
- StreamOV: Streaming Omni-Video Understanding via Evidence-Guided Memory and Response Triggering (NeurIPS 2026 · similarity 0.87)
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.
