Research map / Vision-language and multimodal / Multimodal LLMs

Breaking the Regional Perception Bottleneck of Multimodal Large Language Models via External Reasoning Framework

CVPR 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.

Proposes R-Ground, a multimodal Monte Carlo Tree Search reasoning framework that scales the semantic refinement stage of MLLMs to improve pixel-level referring expression grounding, letting a 7B model match or surpass 72B models.

Problem

MLLMs generating pixel coordinates in natural language show weak gains from parameter and data scaling on grounding tasks, because the semantic refinement stage of the LM decoder is the bottleneck for regional perception.

Approach

Decomposes grounding into a search tree with vision-perception and semantic-refinement actions, uses CLIP-based multimodal alignment scores in the UCT node selection (truncating low-alignment paths), and aggregates leaf candidates via alignment-weighted voting instead of majority voting.

Key results

• R-Ground with Qwen2.5-VL-7B averages 93.16 AP@0.5 across RefCOCO/RefCOCO+/RefCOCOg, versus 90.25 for Qwen2.5-VL-72B and 86.56 for the 7B base model.
• R-Ground with Qwen3-VL-8B reaches 94.47 average AP@0.5, above InternVL3.5-241B (92.39).
• On RefCOCO+, R-Ground (Qwen2.5-VL-7B) improves over Semantic-CoT (88.27 average) to 91.93, and weighted voting beats majority voting (91.93 vs 86.36).

Surprising findings

• Qwen2.5-VL-7B with R-Ground outperforms the 72B Qwen2.5-VL model on grounding benchmarks, despite the 72B model's far greater parameter count.

Limitations

• Relies on multiple MCTS rollouts and depth-4 search, adding inference cost compared with single-pass grounding.
• Alignment scoring depends on CLIP-based matching quality for region and text.
• Evaluation limited to referring expression grounding benchmarks (RefCOCO family).

Experimental setup

Datasets: RefCOCO RefCOCO+ RefCOCOg

Benchmarks: RefCOCO RefCOCO+ RefCOCOg

Metrics: AP@0.5

Baselines: Grounding DINO-L UNINEXT-H ONE-PEACE GLaMM Shikra-7B Ferret-v2-7B InternVL3-8B InternVL3-78B InternVL3.5-241B Qwen2.5-VL 7B Qwen2.5-VL 72B Rex-Thinker-CoT

Builds on: Monte Carlo Tree Search UCT CLIP Chain-of-Thought

Closest papers by meaning

• GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.90)
• DiG: Differential Grounding for Enhancing Fine-Grained Perception in Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
• Grounding Multimodal Large Language Models to the World (ICLR 2024 · similarity 0.90)
• GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.90)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.89)
• Grounded Chain-of-Thought for Multimodal Large Language Models (CVPR 2026 · similarity 0.89)