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

Ruilin Luo; Chufan Shi; Yizhen Zhang; Cheng Yang; Songtao Jiang; Tongkun Guan; Ruizhe Chen; Ruihang Chu

cs.CVMar 4, 2026

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

Authors: Ruilin Luo, Chufan Shi, Yizhen Zhang, Cheng Yang, Songtao Jiang, Tongkun Guan, Ruizhe Chen, Ruihang Chu, +5 more

Organizations: 1Tsinghua University · 3Qwen Team, Alibaba Group · University of South California · University of California San Diego · Zhejiang University · Shanghai Jiao Tong University

Abstract

The cold-start initialization stage plays a pivotal role in training Multimodal Large Reasoning Models (MLRMs), yet its mechanisms remain insufficiently understood. To analyze this stage, we introduce the Visual Attention Score (VAS), an attention-based metric that quantifies how much a model attends to visual tokens. We find that reasoning performance is strongly correlated with VAS (r=0.9616): models with higher VAS achieve substantially stronger multimodal reasoning. Surprisingly, multimodal cold-start fails to elevate VAS, resulting in attention distributions close to the base model, whereas text-only cold-start leads to a clear increase. We term this counter-intuitive phenomenon Lazy Attention Localization. To validate its causal role, we design training-free interventions that directly modulate attention allocation during inference, performance gains of 1 $-$ 2% without any retraining. Building on these insights, we further propose Attention-Guided Visual Anchoring and Reflection (AVAR), a comprehensive cold-start framework that integrates visual-anchored data synthesis, attention-guided objectives, and visual-anchored reward shaping. Applied to Qwen2.5-VL-7B, AVAR achieves an average gain of 7.0% across 7 multimodal reasoning benchmarks. Ablation studies further confirm that each component of AVAR contributes step-wise to the overall gains. The code, data, and models are available at https://github.com/lrlbbzl/Qwen-AVAR.

Figures & tables

Panel of Figure 1 — Figure 1: Analysis of different models’ performance and Visual Attention Score (VAS) distribution. (a) Model performance by mean VAS; (b) VAS distribution across layers.

Figure 2: Performance gains from training-free attention modulation on MathVista, MathVision and MathVerse-VO.

Figure 3: Overview of Visual-Anchored Reflection data synthesis in $3$ steps: Visual Description, Reflection-Enhanced Reasoning, and Visual Anchor Integration.

Configuration	Method Components			Benchmark Performance
Configuration	VARD	AGTO	VARS	MathVista	MathVision	MathVerse-VO	MMStar	MMMU-VAL	MMMU-Pro	Hallusion.	Avg.
Baseline (Qwen2.5-VL-7B)				68.2	25.2	41.1	62.1	58.1	38.3	50.7	49.1
	✓			70.6	32.9	43.5	61.1	55.2	38.7	55.3	51.0
	✓	✓		72.0	34.1	44.0	62.8	58.3	39.8	57.2	52.6
AVAR-Thinker	✓	✓	✓	74.7	37.4	50.4	64.1	63.8	42.9	59.5	56.1

Table 2: Ablation study of our proposed components. Starting from the baseline, we show the performance impact of adding different modules, indicated by a checkmark (✓).

Method	Benchmark Performance							Avg.
Method	MathVista	MathVision	MathVerse-VO	MMStar	MMMU-val	MMMU-Pro	Hallusion.	Avg.
Baseline (Qwen2.5-VL-7B)	68.2	25.2	41.1	62.1	58.1	38.3	50.7	49.1
+ R1-OneVision	63.3	26.3	39.7	54.9	49.9	34.6	43.8	44.6
+ OpenVLThinker	68.9	25.3	37.8	58.7	55.7	36.0	54.1	48.1
+ Vision-SR1	67.6	27.9	42.3	46.9	50.7	36.3	42.1	44.8
+ VARD (Ours)	70.6	32.9	43.5	61.1	55.2	38.7	55.3	51.0

Table 3: Comparison of our Visual-Anchored Reflection Data (VAR) against other data-centric cold-start methods. The best scores are bold ; the second best are underlined .

Model	VAS	Avg. Performance
Qwen2.5-VL-7B	7.5	49.3
Qwen2.5-VL-7B + VARD Data	10.1	51.0
AVAR-CS	13.8	52.6
AVAR-Thinker	18.9	56.1

Table 4: Evolution of VAS and performance across AVAR training stages.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	VARD	AGTO	VARS	MathVista	MathVision	MathVerse-VO	MMStar	MMMU-VAL	MMMU-Pro	HallusionBench	Avg.
Configuration	Method Components			Benchmark Performance
Baseline (Llama-3.2-11B-Vision-Instruct)				48.6	19.7	18.4	49.8	50.7	33.0	40.3	37.2
	✓			56.6	25.5	25.4	58.0	55.2	36.4	45.5	43.2
	✓	✓		57.4	25.2	26.6	58.8	56.2	37.0	46.4	44.0
AVAR-Thinker	✓	✓	✓	61.7	26.9	29.0	61.8	58.6	38.9	50.1	46.7

Appendix

Table 5: Ablation study of our proposed components on Llama-3.2-11B-Vision-Instruct model.

Figure 4: Attention allocation of Qwen2.5-VL-7B, R1-OneVision and OVR-CS on Mathvista.

Figure 5: Attention allocation of ThinkLite-VL, RevisualR1-CS and MIMO-VL-CS on Mathvista.

Figure 6: Attention allocation of Qwen2.5-VL, Vision-R1 and OVR-CS on Mathvision.

Cold-start		RL
Hyperparameter	Value	Hyperparameter	Value
Cutoff length	30,000	Max response length	30,000
Epochs	20	Weight decay	$3\times 10^{-5}$
Learning rate	$5\times 10^{-6}$	Learning rate	$1\times 10^{-6}$
Warm-up ratio	0.1	Warm-up ratio	0.03
Batch size	512	Batch size	256

Appendix

Table 6: Hyperparameter search spaces used in experiments.

Figure 7: AVAR-Thinker on MathVerse-VO: A showcase demonstrating its powerful visual perception and reflective capabilities.

Model Name	Hugging Face Repository
Qwen2.5-VL-7B-Instruct	https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct
R1-OneVision	https://huggingface.co/Fancy-MLLM/R1-Onevision-7B-RL
InternVL2.5-8B	https://huggingface.co/OpenGVLab/InternVL2_5-8B
MM-Eureka	https://huggingface.co/FanqingM/MM-Eureka-Qwen-7B
ThinkLite-VL	https://huggingface.co/russwang/ThinkLite-VL-7B
Revisual-R1-CS	https://huggingface.co/csfufu/Revisual-R1-Coldstart

Appendix

Table 7: List of models and their Hugging Face repositories.

Explore similar work

Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention
May 21, 2026Changyuan Tian, Zhicong Lu, Huaxing Liu +7Multimodal Large Language Models Multimodal Reasoning
AMAP, Alibaba Group · University of Chinese Academy of Sciences · Tsinghua University +1
Mind What Matters for Reasoning: Aligning Cross-Modal Attention via Selective Probability Mass Concentration
Sep 24, 2026Jiaqi Deng, Zonghan Wu, Zhan Heng +3Multimodal Grounding Multimodal Large Language Models
University of Technology Sydney · East China Normal University · The University of New South Wales +2
Deeper Thought, Weaker Aim: Understanding and Mitigating Perceptual Impairment during Reasoning in Multimodal Large Language Models
Mar 15, 2026Ruiying Peng, Xueyu Wu, Jing Lei +3Multimodal Large Language Models VLM Reasoning
Tsinghua Shenzhen International Graduate School · Huawei Technologies

May 21, 2026J/K move · Enter open · S save

Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention

Changyuan Tian, Zhicong Lu, Huaxing Liu, Xiang Wang+6

AMAP, Alibaba Group · University of Chinese Academy of Sciences · Tsinghua University+1

Multimodal Large Language Models Multimodal Reasoning Visually Grounded Reasoning Visual Grounding Faithfulness in VLMs RL for VLMs+1

Open paper AskPDF arXiv

Reinforcement learning with verifiable rewards (RLVR) has emerged as a promising paradigm for advancing complex reasoning in large language models, and recent work extends RLVR to multimodal large language models (MLLMs). This transfer, however, surfaces a faithfulness challenge: faithful perception of task-relevant visual evidence and faithful use of that evidence during reasoning, leading to unsatisfactory gains on multimodal benchmarks. Specifically, existing perception supervision often operates on textual descriptions rather than natively on image regions, and faithful use is largely overlooked, exposing the perception-reasoning disconnect where correctly perceived evidence is dropped or contradicted during reasoning. To close these gaps, we propose Faithful-MR1, a training framework that anchors and reinforces visual attention to address both halves of faithful multimodal reasoning. The Anchoring stage turns perception into an explicit pre-reasoning subtask, supervising a dedicated <Focus> token's attention directly against image regions rather than through textual descriptions. The Reinforcing stage exposes faithful use through counterfactual image intervention, rewarding answer-correct trajectories that concentrate visual attention where vision causally matters. Extensive experiments demonstrate that Faithful-MR1 outperforms recent multimodal reasoning baselines on both Qwen2.5-VL-Instruct 3B and 7B backbones while using substantially less training data.

Abstract

Explore similar work

Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention

Mind What Matters for Reasoning: Aligning Cross-Modal Attention via Selective Probability Mass Concentration

Deeper Thought, Weaker Aim: Understanding and Mitigating Perceptual Impairment during Reasoning in Multimodal Large Language Models