UniCAR-RL: Seeing Better before Thinking Deeper in Visual Mathematics

Yuzhe Li; Hao Yan; Hao Wang; Xingchen Liu; Ya-Qi Yu; Jihao Wu; Minghui Liao; Wei Chen

cs.AISep 12, 2026

UniCAR-RL: Seeing Better before Thinking Deeper in Visual Mathematics

Authors: Yuzhe Li, Hao Yan, Hao Wang, Xingchen Liu, Ya-Qi Yu, Jihao Wu, Minghui Liao, Wei Chen,

Organizations: Huazhong University of Science and Technology · Huawei Inc.

Abstract

Multimodal Large Language Models (MLLMs) often struggle with complex mathematical visual reasoning primarily due to a lack of fine-grained perception, causing initial visual hallucinations to directly trigger cascading reasoning failures. In traditional end-to-end reinforcement learning (RL), sparse rewards fail to decouple perceptual hallucinations from logical missteps, hindering targeted perception optimization. Alternatively, fine-tuning with perception-enhanced CoT data incurs high costs and hallucinations. In this paper, we address these challenges by proposing UniCAR-RL, an annotation-free RL framework. By explicitly decoupling the optimization of perception and reasoning during the training process, it achieves isolation and optimization of both capabilities. Specifically, UniCAR-RL consists of three synergistic branches: 1) a Caption-RL branch that optimizes perception capabilities through verifier-guided reasoning validation; 2) a Reasoning-RL branch that performs logical reasoning based on a gold image description to halt cascading errors; 3) a QA-RL branch that retains native end-to-end alignment to ensure robust question-answering performance. Experiments show that UniCAR-RL substantially improves MLLMs' mathematical and visual reasoning using only raw short-answer data. Furthermore, it demonstrates strong generalization across diverse architectures and scales.

Figures & tables

Figure 1: Motivation of proposed approach. (a) highlights the perception bottleneck of MLLMs in visual mathematical reasoning, but existing RL methods struggle to address it; (b) presents existing perception enhancement methods and their limitations; (c) provides a comparison of error distributions on visual mathematical reasoning tasks.

Figure 2: Overview of UniCAR-RL. It disentangles optimization into three branches: a) the Caption-RL Branch for pure visual perception enhancement, b) the Reasoning-RL Branch for visually-isolated logic optimization, and c) the QA-RL Branch for synergistic end-to-end QA integration.

Close-source Proprietary model
Model	Venue	WeMath	MathVerse	MathVista	MathVision	Avg.
GPT-4o [ 36 ]		68.8	50.2	63.8	30.4	53.4
Claude-3.7-Sonnet [ 37 ]		72.6	52.0	66.8	41.3	58.1
Models based on Qwen3-VL-Instruct
Qwen3-VL-4B [ 38 ]	arXiv 2025	75.0	47.2	73.5	50.1	61.5
Qwen3-VL-8B [ 38 ]	arXiv 2025	79.4	60.8	77.9	52.9	67.8

Table 1: Main results on visual mathematical reasoning benchmarks. The Avg. score is the arithmetic mean over WeMath, MathVerse, MathVista, and MathVision; CVPR-F denotes CVPR Findings. Models marked with ⋆ are evaluated under our setting. Results marked with † are sourced from previous work [ 30 ] . The best value in each column is shown in bold , and the second-best is underlined . For UniCAR-RL ( Verifier ), the Verifier denotes the advanced verifier used during training. This convention applies to all subsequent tables.

Base model ablations (Verifier: Qwen3.5-35B-A3B)
Configuration	WeMath	MathVerse	MathVista	MathVision	Avg.
Qwen2.5-VL-3B (base)	55.9	33.6	59.4	22.5	42.9
+ UniCAR-RL	66.5 (+10.6)	41.8 (+8.2)	68.9 (+9.5)	27.3 (+4.8)	51.1 (+8.2)
Qwen3-VL-4B (base)	75.0	47.2	73.5	50.1	61.5
+ UniCAR-RL	77.8 (+2.8)	50.4 (+3.2)	79.6 (+6.1)	52.9 (+2.8)	65.2 (+3.7)
Verifier ablations

Table 2: Component ablations of UniCAR-RL. We adopt Qwen3-VL-4B (Qwen3.5-35B-A3B) as the default configuration. We then conduct ablations on the base model, parameter scale, and verifier selection. Rows marked (base) denote the pretrained model without UniCAR-RL training and green values report absolute improvement over the corresponding base.

Captioner	Verifier-Solve Accuracy (%) $\uparrow$
Captioner	WeMath	MathVerse	MathVista	MathVision	Avg.
Qwen3-VL-4B (zero-shot)	80.8	62.4	76.6	65.9	71.4
Qwen3-VL-4B + GRPO	82.1 (+1.3)	63.8 (+1.4)	77.8 (+1.2)	67.5 (+1.6)	72.8 (+1.4)
Qwen3-VL-4B + UniCAR-RL	84.9 (+4.1)	65.5 (+3.1)	82.5 (+5.9)	70.8 (+4.9)	75.9 (+4.5)

Table 3: Evaluation results of visual perception. We adopt Gemini 3 Pro as the unified verifier.

Model	ChartQA	HallBench
Models around 4B parameters
Qwen3-VL-4B	83.0	73.7
+UniCAR-RL(Qwen3.5-35B-A3B)	87.9 (+4.9)	75.0 (+1.3)
+UniCAR-RL(Gemini 3 Pro)	89.5 (+6.5)	75.5 (+1.8)
Models around 8B parameters
Qwen3-VL-8B	88.4	75.0

Table 4: Performance comparison on general multimodal benchmarks. HallBench stands for HallusionBench.

Figure 3: Qualitative comparison of mathematical reasoning tasks.

Configuration	WeMath	MathVerse	MathVista	MathVision	Avg.
UniCAR-RL(full)	78.5	51.2	80.4	53.3	65.9
w/o Reasoning-RL Branch	77.2	49.8	78.1	51.4	64.2
w/o Caption-RL Branch (Vanilla GRPO)	76.3	48.5	76.0	50.8	62.8
w/o QA-RL Branch (Base Qwen3-VL-4B)	75.0	47.2	73.5	50.1	61.5

Table 5: Branch ablations of UniCAR-RL. By systematically removing each branch, UniCAR-RL progressively degrades to the standard baseline. The default setup uses Qwen3-VL-4B with the Gemini 3 Pro verifier.

Method	Rollout	Total
Vanilla GRPO (QA-RL)	72.3	297.8
Reasoning-RL	68.4	292.7
Caption-RL	305.1	525.5
UniCAR-RL iteration (sum)	445.8	1116.0 ( $\sim$ 3.75 $\times$ )

Table 6: Training Cost Analysis of UniCAR-RL.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Venue	Backbone	Size	Train	Distill.	Percep.
Reinforcement Learning for Mathematical Reasoning
SRPO [ 19 ]	NeurIPS’25	Qwen2.5-VL	7B/32B	SFT+RL	$\checkmark$	$\checkmark$
Vision-R1 [ 21 ]	ICLR’26	Qwen2.5-VL	7B/32B/72B	SFT+RL	$\checkmark$	$\checkmark$
MM-Eureka [ 23 ]	TMLR’26	Qwen2.5-VL	7B/32B	Zero-RL	$\times$	$\checkmark$
R1-VL [ 24 ]	ICCV’25	Qwen2-VL	7B	SFT+RL	$\times$	$\checkmark$
URSA [ 22 ]	NeurIPS’25	Qwen2.5-Math	8B	SFT+RL	$\checkmark$	$\checkmark$

Appendix

Table 7: Comparison of representative methods for multimodal mathematical reasoning. “Train” indicates the training paradigm, where “Zero-RL” denotes direct RL optimization from a pretrained or instruction-tuned model without task-specific cold-start SFT, “SFT+RL” denotes a cold-start supervised stage followed by reinforcement learning, and “–” indicates non-RL methods. “Distill.” indicates whether the method explicitly relies on distilled reasoning traces, teacher-generated supervision, or synthetic high-quality CoT/QA data. “Percep.” denotes whether the method explicitly enhances visual perception, grounding, diagram understanding, or visual reasoning capabilities.

Model	ChartQA	HallBench	V*Bench	HR-Bench 4K	HR-Bench 8K
Qwen3-VL-4B	83.0	73.7	86.4	80.0	73.6
+UniCAR-RL	89.5	75.5	88.5	81.6	76.4
Qwen3-VL-8B	88.4	75.0	88.7	81.2	76.5
+UniCAR-RL	90.6	75.4	90.1	82.8	78.3

Appendix

Table 8: Additional perception results on ChartQA, HallBench, V*Bench and HR-Bench. Gemini 3 Pro is used as the verifier during training by default.

Panel of Figure 8 — Figure 8: Task-specific prompt templates and output evaluation protocols. Placeholders enclosed in braces are replaced with the corresponding benchmark questions and answer options at inference time.

VLM Reasoning

Vision-language model reasoning studies how multimodal systems use visual evidence and language to answer questions, plan, and explain.

See all 853 papers

Multimodal Mathematical Reasoning with Diverse Solving Perspective
Jul 3, 2025Wenhao Shi, Zhiqiang Hu, Yi Bin +4Multimodal Reasoning Multimodal CoT Reasoning
Tongji University, Shanghai, China · National University of Singapore, Singapore · Tencent, Beijing, China +2
MathVis-Fine: Aligning Visual Supervision with Necessity via Progressive Dependency-Guided Training for Multimodal Mathematical Reasoning
Jun 16, 2026Wanshi Xu, Haokun Zhao, Haidong Yuan +2Multimodal Reasoning Multimodal Reward Modeling
School of ECE, Peking University · College of Computer Science and Artificial Intelligence, Fudan University · School of Software and Microelectronics, Peking University +1
Func-R1: Incentivizing Mathematical Function Reasoning in Multimodal Large Language Models
Sep 13, 2026Mingze Yin, Xiaohan Wang, Dian Li +8Multimodal Large Language Models Mathematical Reasoning
College of Computer Science and Technology, Zhejiang University · Tencent 3AI Thrust, HKUST (GZ) · Zhongguancun Academy +3

Jul 3, 2025J/K move · Enter open · S save

Multimodal Mathematical Reasoning with Diverse Solving Perspective

Wenhao Shi, Zhiqiang Hu, Yi Bin, Guoqing Wang+3

Tongji University, Shanghai, China · National University of Singapore, Singapore · Tencent, Beijing, China+2

Multimodal Reasoning Multimodal CoT Reasoning RL for VLMs

Open paper AskPDF arXiv

Recent progress in large-scale reinforcement learning (RL) has notably enhanced the reasoning capabilities of large language models (LLMs), especially in mathematical domains. However, current multimodal LLMs (MLLMs) for mathematical reasoning often rely on one-to-one image-text pairs and single-solution supervision, overlooking the diversity of valid reasoning perspectives and internal reflections. In this work, we introduce MathV-DP, a novel dataset that captures multiple diverse solution trajectories for each image-question pair, fostering richer reasoning supervision. We further propose Qwen-VL-DP, a model built upon Qwen-VL, fine-tuned with supervised learning and enhanced via group relative policy optimization (GRPO), a rule-based RL approach that integrates correctness discrimination and diversity-aware reward functions. Our method emphasizes learning from varied reasoning perspectives and distinguishing between correct yet distinct solutions. Extensive experiments on the MathVista's minitest and Math-V benchmarks demonstrate that Qwen-VL-DP significantly outperforms prior base MLLMs in both accuracy and generative diversity, highlighting the importance of incorporating diverse perspectives and reflective reasoning in multimodal mathematical reasoning.

Abstract

VLM Reasoning

Explore similar work

Multimodal Mathematical Reasoning with Diverse Solving Perspective

MathVis-Fine: Aligning Visual Supervision with Necessity via Progressive Dependency-Guided Training for Multimodal Mathematical Reasoning

Func-R1: Incentivizing Mathematical Function Reasoning in Multimodal Large Language Models