Research map / Vision-language and multimodal / Multimodal LLMs

DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

Read the paperCheck your idea against 108,967 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.

Trains Qwen2.5-VL-7B end-to-end with GRPO and an outcome-based reward (with a conditional tool bonus) so it learns to zoom into image regions during reasoning ("thinking with images") without SFT cold start.

Problem

Vision-language models reason mostly in text and struggle to integrate fine-grained visual evidence, especially on high-resolution images; prior visual-reasoning pipelines rely on hand-designed workflows or SFT on reasoning data.

Approach

Encapsulate the VLM's native grounding ability as a crop/zoom-in tool invoked within an interleaved multimodal CoT, trained purely by RL with outcome rewards; a data selection filter (difficulty, verification, perception-utility) and a reward bonus given only for correct answers that used active perception incentivize effective zoom-in behavior.

Key results

• DeepEyes-7B reaches 90.1 on V* Bench (+18.9 over Qwen2.5-VL-7B) and improves HR-Bench 8K by 7.3 points (72.6).
• On MME-RealWorld-Lite, DeepEyes-7B scores 53.2 overall, above Qwen2.5-VL-32B (45.6).
• Ablation: iMCoT reaches 72.6 on HR-8K versus 60.8 for RL with text-only CoT; the conditional tool reward outperforms unconditional and no bonus.

Surprising findings

• Without the conditional tool reward the model quickly stops using zoom-in actions, and an unconditional bonus leaves engagement minimal and static.
• Text-only RL reaches higher V* accuracy (88.5) than iMCoT on HR-4K (75.4 vs 75.1), so iMCoT is not uniformly better.

Limitations

• Only crop tool implemented; no search or drawing tools.
• Relies on a relatively weak 7B base model, leading to shortcuts and inaccurate localization.
• Grounding drift and reasoning errors cause failures on some cases.

Experimental setup

Datasets: V* training set COCO2017 ArxivQA ThinkLite-VL

Benchmarks: V* Bench HR-Bench 4K HR-Bench 8K MME-RealWorld-Lite refCOCO refCOCO+ refCOCOg ReasonSeg POPE MathVista MathVerse MathVision

Metrics: accuracy IoU overall score

Baselines: GPT-4o o3 LLaVA-OneVision Qwen2.5-VL SEAL DyFo ZoomEye Pixel-Reasoner RL w. Text-only CoT

Builds on: GRPO Qwen2.5-VL V* benchmark training data ThinkLite-VL

Closest papers by meaning

• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.90)
• SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.90)
• CoFFT: Chain of Foresight-Focus Thought for Visual Language Models (NeurIPS 2025 · similarity 0.90)
• Mind the GAP: Glimpse-based Active Perception improves generalization and sample efficiency of visual reasoning (ICLR 2025 · similarity 0.90)
• See Further, Think Deeper: Advancing VLM's Reasoning Ability with Low-level Visual Cues and Reflection (CVPR 2026 · similarity 0.90)