Research map / Video / Video-language understanding
APPO: Attention-guided Perception Policy Optimization for Video Reasoning
Henghui Du, Chang Zhou, Xi Chen, Di Hu
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 |
|---|---|---|---|
| replace GRPO/DAPO with APPO (attention-guided token re-weighting) | Qwen2.5-VL-3B · SEED-Bench-R1 L3 (OOD) | accuracy: 31.8 -> 35.0 (vs DAPO) | Improved |
| replace DAPO with APPO | Qwen2.5-VL-7B · SEED-Bench-R1 L1 (in-dist) | accuracy: 50.0 -> 50.5 | Improved |
| replace DAPO with APPO | Qwen2.5-VL-3B · NExT-GQA | mIoU: 10.6 -> 11.1 | Improved |
| replace DAPO with APPO | Qwen2.5-VL-3B · MVBench | accuracy: 62.8 -> 62.4 | No clear effect |
| replace DAPO with APPO | Qwen2.5-VL-7B · NExT-QA | accuracy: 79.9 -> 79.6 | No clear effect |
| RL (GRPO) instead of SFT | Qwen2.5-VL-7B · SEED-Bench-R1 L1 (in-dist) | accuracy: 40.2 -> 49.0 | Improved |
| enhance perception model from Qwen2.5-VL-7B to 32B (ablation) | Qwen3-8B reasoner · SEED-Bench-R1 L1 | accuracy: +1.4% | Improved |
| enhance reasoning model from Qwen3-8B to OpenAI-o3 (ablation) | Qwen2.5-VL-7B perception · SEED-Bench-R1 L1 | accuracy: +0.7% | No clear effect |
| increase K1 (tokens per response for attention aggregation) (ablation) | APPO · SEED-Bench-R1 (Qwen2.5-VL-7B) | accuracy | Improved |
| increase K2 beyond 3 (selected frames per response) (ablation) | APPO · SEED-Bench-R1 (Qwen2.5-VL-7B) | accuracy | Hurt |
| increase K3 (perception tokens per response) to large values (ablation) | APPO · SEED-Bench-R1 (Qwen2.5-VL-7B) | accuracy | Hurt |
| vary token weight scale alpha from 1.2 to 2.0 (ablation) | APPO · SEED-Bench-R1 (Qwen2.5-VL-7B) | accuracy | Mixed |
Closest papers by meaning
- Act2See: Emergent Active Visual Perception for Video Reasoning (CVPR 2026 · similarity 0.89)
- Beyond Perceptual Shortcuts: Causal-Inspired Debiasing Optimization for Generalizable Video Reasoning in Lightweight MLLMs (CVPR 2026 · similarity 0.88)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.87)
- NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation (NeurIPS 2025 · similarity 0.87)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.87)
- Let VLMs Grade Their Own Thoughts: A Self-Quantification Approach to Reasoning-Aware Reward Modeling (CVPR 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.
