Research map / Vision-language and multimodal / Multimodal LLMs
Learning Where to Look: Observation Policy Optimization for Thinking with Images
Junfeng Wang, Jiawei Liu, Yongchao Xu, Tao Jiang, jiangbo Ai, Jin Zhang, Chong Wang, Qixing Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes observation policy optimization to train vision-language models to choose better zoom/crop observation actions when thinking with images, motivated by error analysis showing many failures stem from incorrect observations.
Problem
Complex visual question answering with high-resolution images, where vision-language models must actively acquire local visual evidence by zooming into regions; many failures come from incorrect observation actions (failing to localize key regions or obtaining insufficient crops).
Approach
Treats the choice of where to look (observation actions such as crops) as a learnable policy and optimizes it directly, rather than optimizing only the final answer; details beyond the abstract are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with OPO instead of base Qwen3-VL-2B with zoom tool | Qwen3-VL-2B · V* | overall accuracy: 75.9 -> 85.9 | Improved |
| train with OPO instead of base Qwen3-VL-8B | Qwen3-VL-8B · HR-Bench 8K | overall accuracy: 76.5 -> 80.8 | Improved |
| OPO instead of DeepEyes-style outcome GRPO | Qwen3-VL-2B · HR-Bench 8K | overall accuracy: 71.1 -> 76.4 | Improved |
| OPO instead of DeepEyes-style GRPO | Qwen2.5-VL-7B · HR-Bench 8K | overall accuracy: 68.0 -> 68.1 | No clear effect |
| OPO instead of DeepEyes-style GRPO on V* spatial | Qwen3-VL-8B · V* Spatial | accuracy: 93.4 -> 90.8 | Hurt |
| add UOB and ELAA (vs plain outcome-level GRPO) (ablation) | Qwen3-VL-2B · HR-Bench 4K | overall accuracy: 71.5 -> 78.4 | Improved |
| add ELAA on top of UOB branching (ablation) | Qwen3-VL-2B · HR-Bench 8K | overall accuracy: 73.8 -> 76.4 | Improved |
| branch on UOB without ELAA vs no branching (ablation) | Qwen3-VL-2B · V* | overall accuracy: 82.7 -> 83.8 | Improved |
| use uncertainty-guided event selection instead of random branching (ablation) | Qwen3-VL-2B · HR-Bench 8K | overall accuracy: 73.0 -> 76.4 | Improved |
| add crop redundancy filtering (ablation) | Qwen3-VL-2B · V* | overall accuracy: 84.8 -> 85.9 | Improved |
| rollout budget M=4,B=1 instead of M=8,B=2 (ablation) | Qwen3-VL-2B · HR-Bench 8K | overall accuracy: 76.4 -> 72.5 | Hurt |
| base trajectories B=4 instead of B=2 (M=8) (ablation) | Qwen3-VL-2B · HR-Bench 4K | overall accuracy: 78.4 -> 77.1 | Hurt |
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.
- DeepEyes — DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) · better on 1 of 1 shared result
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
- Diversity Over Frequency: Rethinking Tool Use in Visual Chain-of-Thought Agents (ICML 2026 · similarity 0.90)
- The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design (ICML 2026 · similarity 0.90)
- When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.89)
- V-ABS: Action-Observer Driven Beam Search for Dynamic Visual Reasoning (ICML 2026 · similarity 0.89)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.89)
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.89)
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.
