Research map / Vision-language and multimodal / Multimodal LLMs

DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal 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 DeepScan, a training-free bottom-up framework (Hierarchical Scanning, Refocusing, Evidence-Enhanced Reasoning) that couples LVLMs with a search expert (BLIP-ITM) and visual expert (LangSAM) to improve visually grounded reasoning on fine-grained benchmarks.

Problem

LVLMs struggle to localize small, task-relevant visual evidence in high-resolution images; existing one-shot top-down grounding is distracted by attention sink/drift and semantically similar objects.

Approach

Bottom-up grounding: scan image patches for local cues using attention maps, convert cues to interior point proxies, segment evidence with a point-prompted visual expert, then refocus the crop via zoom-in/zoom-out search judged by the LVLM and answer from a hybrid memory of fine and coarse views.

Key results

• DeepScan with Qwen2.5-VL-7B reaches 90.6% overall on V* Bench, a 16.3-point gain over the 74.3% baseline and above DeepEyes (90.0%).
• On TreeBench, DeepScan reaches 42.5 overall and 37.3 mIoU, improving Qwen2.5-VL-7B by 5.5 points.
• DeepScan-72B reaches 94.2% on V* Bench versus 84.8% for Qwen2.5-VL-72B.

Surprising findings

• RL-based visually grounded methods gain little on second-order reasoning tasks, and the authors hypothesize RL mainly biases LVLMs toward perception.
• Excessive zoom-in can remove essential context and degrade LVLM reasoning, so more aggressive grounding is not always better.

Limitations

• Higher inference latency than one-shot evidence-detection pipelines.
• Fixed small patch size is inefficient for simple images with salient evidence.
• Merged minimal enclosing crop introduces inter-evidence noise for widely separated spatial-relation questions.

Experimental setup

Datasets: V* Bench HR-Bench TreeBench

Benchmarks: V* Bench HR-Bench 4K HR-Bench 8K TreeBench

Metrics: multiple-choice accuracy mIoU latency

Baselines: DeepEyes PixelReasoner TreeVGR Thyme-VL ZoomRefine Dyfo SEAL GPT-4o LLaVA-OneVision-7B InternVL3-8B

Builds on: DyFo ZoomRefine LangSAM GradCAM Otsu's method

Closest papers by meaning

• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.90)
• Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Method (ICLR 2026 · similarity 0.90)
• VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning (ICML 2026 · similarity 0.90)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.90)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.90)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.90)