Research map / Vision-language and multimodal / Visual question answering
Learning to Focus and Precise Cropping:A Reinforcement Learning Framework with Information Gaps and Grounding Loss for MLLMs
Xuanpu Zhao, Zhentao Tan, Dianmo Sheng, Tianxiang Chen, Yao Liu, Yue Wu, Tao Gong, Qi Chu, Nenghai Yu
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 information gap (downsampled global image, full-res crop) in stage I (ablation) | Qwen2.5-VL-7B RL agent · HR-Bench 8K (1,024 tokens) | overall accuracy: 61.0 -> 69.9 | Improved |
| add information gap in stage I (ablation) | Qwen2.5-VL-7B RL agent · V* (1,024 tokens) | overall accuracy: 72.3 -> 78.0 | Improved |
| add grounding reward stage II with VisualProbe data and L1 reward (ablation) | Qwen2.5-VL-7B RL agent · HR-Bench 8K (1,024 tokens) | overall accuracy: 69.9 -> 72.1 | Improved |
| stage II trained on first-stage source data instead of VisualProbe (ablation) | Qwen2.5-VL-7B RL agent · HR-Bench 8K (1,024 tokens) | overall accuracy: 69.9 -> 68.1 | Hurt |
| add L1 distance reward on top of VisualProbe data (ablation) | Qwen2.5-VL-7B RL agent · V* (1,024 tokens) | overall accuracy: 76.4 -> 80.6 | Improved |
| add L1 distance reward on top of VisualProbe data (ablation) | Qwen2.5-VL-7B RL agent · HR-Bench 8K (1,024 tokens) | overall accuracy: 72.7 -> 72.1 | No clear effect |
| replace crop with GT or noise to test crop reliance (information gap training) (ablation) | Qwen2.5-VL-7B RL agent · HR-Bench 8K | accuracy delta: 4.5 -> 18.5 | Improved |
| use 1,024-token input vs 16,384 for competitors, time per question | Qwen2.5-VL-7B RL agent · HR-Bench 8K | seconds per question: 12.4 (DeepEyes) -> 2.8 | Improved |
| replace DeepEyes crop with ground truth box or noise (preliminary) | Qwen2.5-VL-7B RL agent · HR-Bench 8K (16,384 tokens) | overall accuracy: 68.0 vs 69.5 GT vs 68.5 noise | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DeepEyes — DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) · better on 6 of 6 shared results
- Pixel Reasoner — Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (NeurIPS 2025) · better on 3 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
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.91)
- FineRS: Fine-grained Reasoning and Segmentation of Small Objects with Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- GRIT: Teaching MLLMs to Think with Images (NeurIPS 2025 · similarity 0.88)
- Cross-modal Information Flow in Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
- GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.87)
- iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning (ICML 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.
