Research map / Vision-language and multimodal / Multimodal LLMs

VCU-Bridge: Hierarchical Visual Connotation Understanding via Semantic Bridging

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.

Introduces VCU-Bridge, a three-level (perception, semantic bridge, connotation) framework and HVCU-Bench, then uses an MCTS-driven hierarchical data pipeline to instruction-tune Qwen3-VL-4B, improving connotation reasoning and general benchmarks.

Problem

MLLMs treat low-level perception and high-level connotative reasoning in isolation, and existing benchmarks decouple these levels, so they cannot diagnose where hierarchical visual reasoning breaks down.

Approach

Model visual understanding as a hierarchy of perception, semantic bridge, and connotation QA chains linked by support constraints; evaluate per-level and full-chain accuracy; and generate training chains with MCTS plus level-wise validation.

Key results

• Qwen3-VL-4B-Bridge improves HVCU-Bench score by +6.17% over its base model and gains +7.26% average on MMStar, with +2.53% average across four general benchmarks.
• GPT-4o reaches 52.24 overall score on HVCU-Bench versus 87.18 for humans, with a -23.75% gap at the connotation level on Implication Understanding.
• Providing lower-level context raises GPT-4o's overall score from 52.24 to 68.18 (+15.94%).

Surprising findings

• Training only on connotation-level data (L3-only) is weaker than hierarchical data, and adding lower-level supervision improves results even with fewer complete chains.
• Fine-tuning on Implication-only data transfers gains to Aesthetic and Affective tasks that received no direct training signal.

Limitations

• Benchmark is small (1,050 samples) and relies on Gemini-2.5-Pro generation with human validation.
• Human baseline uses undergraduate students with three annotators per level.

Experimental setup

Datasets: II-Bench EEmo-Bench LayerD

Benchmarks: HVCU-Bench MMBench MMMU MMStar HallusionBench

Metrics: accuracy per level full-chain accuracy overall score

Baselines: GPT-4o Qwen3-VL-8B-Instruct Qwen3-VL-4B-Instruct LLaVA-1.6 Gemma3 InternVL3.5 Phi-4-Multimodal-Instruct Phi-3.5-Vision-Instruct Deepseek-VL2 Qwen3-VL-4B-Direct

Builds on: Monte Carlo Tree Search ReST-MCTS* Socratic-MCTS LLaMA-Factory

Closest papers by meaning

• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.90)
• Unveiling Visual Perception in Language Models: An Attention Head Analysis Approach (CVPR 2025 · similarity 0.89)
• JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (NeurIPS 2024 · similarity 0.89)
• Visual Thoughts: A Unified Perspective of Understanding Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.89)
• VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models (ICLR 2026 · similarity 0.89)