Research map / Vision-language and multimodal / Multimodal LLMs

Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Method

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

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.

Introduces TreeBench, a 405-question expert-annotated benchmark with traceable bounding-box evidence for "thinking with images", and TreeVGR, an RL method (GRPO with dual precision/recall IoU rewards) on Qwen2.5-VL-7B that improves grounded reasoning.

Problem

Existing benchmarks do not holistically evaluate visual grounded reasoning (thinking with images): they lack traceable evidence (box-level localization), fine-grained small-target perception, and second-order reasoning. Existing RL methods supervise only the final answer, not intermediate localizati

Approach

A benchmark whose questions come with ground-truth boxes for every target, scored on both answer accuracy and box mIoU, plus an RL reward that combines accuracy, format, and a dual IoU (recall term to cover all targets, precision term to discourage enumerated empty boxes) on top of a cold-start SFT stage with boxes in text form.

Key results

• TreeVGR-7B reaches 50.4 overall and 44.0 mIoU on TreeBench, +13.4 over Qwen2.5-VL-7B, while o3-0416 scores 54.8.
• TreeVGR-7B reaches 91.1 on V* Bench (+16.8 over Qwen2.5-VL-7B) and 54.9 on MME-RealWorld-Lite (+12.6).
• Removing the precision IoU term makes the model enumerate boxes and fail to output answers (0.0 accuracy on TreeBench).

Surprising findings

• Without the precision term the policy learns to enumerate boxes to inflate recall, causing it to stop producing final answers.
• Reasoning-category performance correlates weakly with localization mIoU, unlike perception performance.

Limitations

• Limited to a 7B base model (Qwen2.5-VL-7B), so scalability is untested.
• TreeBench contains only 405 QA pairs.
• TreeVGR is not designed for long-form math, science, or coding reasoning.

Experimental setup

Datasets: SA-1B VGR-158K V* training set VisDrone

Benchmarks: TreeBench V* Bench HR-Bench-4K HR-Bench-8K MME-RealWorld-Lite CV-Bench MMVP MMBench POPE AI2D ChartQA

Metrics: accuracy mIoU

Baselines: GPT-4o-1120 o3-0416 Gemini-2.5-Flash-0520 Gemini-2.5-Pro-0605 Qwen2.5-VL-7B Qwen2.5-VL-72B InternVL3-8B InternVL3-78B LLaVA-OneVision-7B LLaVA-OneVision-72B DeepEyes-7B Pixel-Reasoner-7B

Builds on: GRPO DeepEyes Qwen2.5-VL VGR-158K V*

Closest papers by meaning

• DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.90)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.88)
• Evidence-RL: Towards Evidence-intensive Visual Reasoning (NeurIPS 2026 · similarity 0.88)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.88)
• RegionReasoner: Region-Grounded Multi-Round Visual Reasoning (ICLR 2026 · similarity 0.87)
• No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.87)