Research map / Vision-language and multimodal / Multimodal LLMs

Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

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.

Proposes Region-to-Image Distillation, which synthesizes VQA on micro-crops with teacher consensus and distills them back to full images with box overlays, then trains Qwen-VL students with DAPO and introduces the ZoomBench benchmark.

Problem

MLLMs struggle with fine-grained perception when decisive evidence is small and overwhelmed by global context; existing "Thinking-with-Images" zoom agents fix this at inference time but incur high latency from repeated tool calls and re-encoding.

Approach

Use zooming only at training time: generate perception questions and consensus answers on micro-crops with strong teachers, then map them back to the full image using bounding-box overlays and a spatial constraint prompt, so a single forward pass on the full image internalizes zooming benefits.

Key results

• ZwZ-8B improves Qwen3-VL-8B average from 62.86 to 68.12 and ZoomBench accuracy from 37.87 to 58.11.
• ZwZ-8B reduces the global-regional zooming gap on ZoomBench to 15.26 points versus 25.21 for Qwen3-VL-8B.
• Training Qwen3-VL-8B on 10K R2I samples gives 67.37 average versus 63.71 for MM-Self-Instruct (65K) and 63.24 for Oasis (500K).

Surprising findings

• Many agentic Thinking-with-Images models (DeepEyes, Thyme) perform nearly as well in direct-answer mode without any tool calls.
• Bounding-box overlay on the image gives better results than providing box coordinates in text or no box at all during distillation.

Limitations

• Spatial reasoning and multi-object perception are not covered, so TreeBench is not evaluated.
• Counting remains hard even with zooming, per dual-view analysis.
• Relies on strong teacher models and consensus filtering, which adds data-synthesis cost.

Experimental setup

Datasets: SA-1B LAION MetaCLIP Visual Genome CC12M STPLS3D DeepEyes data Thyme-RL data TreeVGR-RL data Oasis MM-Self-Instruct

Benchmarks: ZoomBench HR-Bench-4K HR-Bench-8K V*Bench CV-Bench MME-RealWorld CountQA ColorBench MMStar BabyVision LOKI FakeCLUE

Metrics: accuracy average accuracy zooming gap relative-attention coverage inference speed

Baselines: GPT-5.1 Gemini-3-Flash Qwen3-VL-235B GLM-4.5V Kimi-K2.5 MiniCPM-4.5-V MiMo-VL-7B-RL DeepEyes DeepEyesV2 Thyme Mini-o3 Pixel-Reasoner

Builds on: DAPO DeepEyes Thyme Qwen3-VL SAM3

Closest papers by meaning

• Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation (NeurIPS 2026 · similarity 0.92)
• Seeing What Matters: A Training-Free Self-Guided Framework for Multimodal Detail Perception and Reasoning (CVPR 2026 · similarity 0.91)
• Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception (ICLR 2026 · similarity 0.91)
• Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models (AAAI 2025 · similarity 0.90)
• CoVFT: Context-aware Visual Fine-tuning for Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
• Zooming from Context to Cue: Hierarchical Preference Optimization for Multi-Image MLLMs (NeurIPS 2025 · similarity 0.90)