Research map / Vision-language and multimodal / Multimodal LLMs

iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 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 iVGR, an RL dual-stream method that trains a textual CoT stream to match high-quality grounded CoT references via an LLM-judged consistency reward, so visual localization is internalized without explicit boxes at inference.

Problem

Explicit bounding-box grounding inside chain-of-thought often degrades MLLM fine-grained VQA accuracy relative to plain textual CoT, and the benefit of visually grounded reasoning at inference is unclear.

Approach

Train a policy with two rollout streams per query: a grounded stream with box predictions (box IoU reward) and a textual stream rewarded for semantic consistency with the best grounded rollout kept in a per-query archive, scored by an LLM judge; at inference only textual CoT is used, optionally with crop tools.

Key results

• iVGR-Qwen2.5-VL-7B reaches 76.6 average on fine-grained and general VQA vs 70.1 for Qwen2.5-VL-7B and 74.7 for TreeVGR-7B, with V* 86.4 and HR8K 75.5.
• iVGR-Qwen3-VL-8B averages 80.6 vs 76.2 for Qwen3-VL-8B; iVGR-Qwen3-VL-32B averages 82.9 vs 79.6.
• Ablation: consistency reward raises the Qwen2.5-VL-7B variant from 71.0 to 71.5 average, and rollout archive further to 72.4 on the ablation subset.

Surprising findings

• Off-the-shelf grounded CoT models (DeepEyes, TreeVGR) often perform worse with explicit boxes than with plain textual CoT at inference.
• Training only on the grounded stream and evaluating with textual CoT outperforms training only on the textual stream.

Limitations

• Requires bounding-box annotations for training data.
• Depends on an external LLM judge whose hallucinations can corrupt the consistency reward.
• No crop tool used during training, so grounded references may miss details in very high-resolution images.

Experimental setup

Datasets: TreeVGR-SFT-35K TreeVGR-RL-37K V* VisDrone OpenMMReasoner ArxivQA

Benchmarks: V* HR4K HR8K MME-RW-Lite POPE RealWorldQA CV-Bench-2D CV-Bench-3D ChartQA AI2D WeMath MMStar

Metrics: accuracy IoU

Baselines: TreeVGR-7B DeepEyes-7B Thyme-7B GRIT-3B PixelReasoner-7B Qwen2.5-VL-7B Qwen3-VL-8B Qwen3-VL-32B

Builds on: GRPO TreeVGR DeepEyes Qwen2.5-VL

Closest papers by meaning

• Grounded Reinforcement Learning for Visual Reasoning (NeurIPS 2025 · similarity 0.91)
• GRIT: Teaching MLLMs to Think with Images (NeurIPS 2025 · similarity 0.91)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.89)
• Start Small, Think Big: Curriculum-based Relative Policy Optimization for Visual Grounding (AAAI 2026 · similarity 0.89)
• Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (CVPR 2026 · similarity 0.89)
• Grounded Chain-of-Thought for Multimodal Large Language Models (CVPR 2026 · similarity 0.89)