Research map / Vision-language and multimodal / Multimodal LLMs

From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning

ICLR 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 Visual Attention Score (VAS) to show MLRM reasoning correlates with visual attention, identifies "Lazy Attention Localization" in cold-start, and proposes AVAR (visual-anchored data, attention-guided losses, visual reward) achieving +7.0% average over Qwen2.5-VL-7B.

Problem

Multimodal cold-start (SFT initialization before RL) for multimodal large reasoning models gives only marginal gains versus text-only cold-start, with no clear mechanistic explanation.

Approach

Quantify visual reliance via an attention-based Visual Attention Score, observe that multimodal cold-start fails to raise it while text-only cold-start does, and reshape attention during cold-start and RL via visual-anchored data synthesis, attention-guided losses, and an attention-based reward.

Key results

• AVAR-Thinker (Qwen2.5-VL-7B based) averages 56.1 across 7 benchmarks vs 49.1 baseline (+7.0%), with MathVision +12.2 and HallusionBench +8.8.
• VAS correlates with reasoning performance at Pearson r=0.9616 across representative 7B MLRMs.
• Training-free attention modulation yields 1-2% gains on Qwen2.5-VL-7B, Revisual-R1-CS, and OVR-CS.

Surprising findings

• Multimodal cold-start (e.g., R1-OneVision) leaves visual attention near the base model, while text-only cold-start raises it (Lazy Attention Localization).
• R1-OneVision data reduces average performance below baseline when used for cold-start (44.6 vs 49.1 in the data comparison).

Limitations

• Correlation analysis across a small set of existing models (about 10) may not establish causality; causal evidence is limited to training-free interventions.
• Training-free modulation gains are modest (1-2%) and hyperparameters were chosen on the evaluation setting.
• Heavy reliance on closed-source models (Gemini 2.5-Pro) for data synthesis.

Experimental setup

Datasets: R1-ShareVL Geo3K M3COT AlgoPuzzleVQA SOLIDGEO Super-CLEVER AI2D MathVista

Benchmarks: MathVista MathVision MathVerse-VO MMMU-VAL MMMU-Pro MMStar HallusionBench

Metrics: accuracy average accuracy Pearson correlation (VAS vs performance)

Baselines: Qwen2.5-VL-7B R1-OneVision OpenVLThinker ThinkLite-VL MM-Eureka-7B Vision-R1 VLAA-Thinker-7B Vision-SR1 Revisual-R1 OVR

Builds on: GRPO DeepSeek-R1 LlamaFactory VeRL

Closest papers by meaning

• Look-Back: Implicit Visual Re-focusing in MLLM Reasoning (AAAI 2026 · similarity 0.91)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.91)
• Look on Demand: A Cognitive Scheduling Framework for Visual Evidence Acquisition in Multimodal Reasoning (ICML 2026 · similarity 0.91)
• Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning (NeurIPS 2025 · similarity 0.90)
• VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning (ICML 2026 · similarity 0.90)
• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.90)