M$^3$-ACE: Rectifying Visual Perception in Multimodal Math Reasoning via Multi-Agentic Context Engineering

Peijin Xie; Zhen Xu; Bingquan Liu; Baoxun Wang

cs.AIMar 9, 2026

M $^3$ -ACE: Rectifying Visual Perception in Multimodal Math Reasoning via Multi-Agentic Context Engineering

Authors: Peijin Xie, Zhen Xu, Bingquan Liu, Baoxun Wang

Organizations: Platform and Content Group Tencent

Abstract

Multimodal large language models have recently shown promising progress in visual mathematical reasoning. However, their performance is often limited by a critical yet underexplored bottleneck: inaccurate visual perception. Through systematic analysis, we find that the most failures originate from incorrect or incomplete visual evidence extraction rather than deficiencies in reasoning capability. Moreover, models tend to remain overly confident in their initial perceptions, making standard strategies such as prompt engineering, multi-round self-reflection, or posterior guidance insufficient to reliably correct errors. To address this limitation, we propose M3-ACE, a multi-agentic context engineering framework designed to rectify visual perception in multimodal math reasoning. Instead of directly aggregating final answers, our approach decouples perception and reasoning by dynamically maintaining a shared context centered on visual evidence lists. Multiple agents collaboratively contribute complementary observations, enabling the system to expose inconsistencies and recover missing perceptual information. To support stable multi-turn collaboration, we further introduce two lightweight tools: a Summary Tool that organizes evidence from different agents into consistent, complementary, and conflicting components, and a Refine Tool that filters unreliable samples and guides iterative correction. Extensive experiments demonstrate that M3-ACE substantially improves visual mathematical reasoning performance across multiple benchmarks. Our method establishes new state-of-the-art results 89.1 on the MathVision benchmark and achieves consistent improvements on other related datasets, including MathVista and MathVerse. These results highlight the importance of perception-centric multi-agent collaboration for advancing multimodal reasoning systems.

Figures & tables

Figure 1: Visual Evidence as the Bottleneck — Decoupling Perception and Reasoning.

Figure 2: Diagnosing the Visual Evidence Bottleneck.

		Gemini 25 pro		GPT-5
All 304 Samples		SP steps	VE list	SP steps	VE list
	Total	208 (68.4%)		210 (69.1%)
	✓	205 (98.6%)	170 (81.7%)	201 (95.7%)	170 (80.9%)
Success	✘	3 (1.4%)	38 (18.3%)	9 (4.3%)	40 (19.1%)
	Total	96 (31.6%)		94 (30.9%)
	✓	67 (69.8%)	12 (12.5%)	62 (66.0%)	16 (17.0%)

Table 1: Decoupling Trajectory and Visual Evidence.

Add Info	Gemini 2.5 pro		GPT 5
Add Info	VE	Answer	VE	Answer
GT judge	22/96	29/96	25/94	31/94
GT judge + GT answer	22/84	-	22/78	-
GT VE	-	88/96	-	89/94

Table 2: Self-Correction under Idealized Supervision

Reflect Mode	Gemini2.5 pro		GPT 5
Reflect Mode	VE	Answer	VE	Answer
Orignal	182	208	186	210
1st turn prompt engineering	181 -1	212 +4	190 +4	207 -3
2nd turn reflect on answer	189 +7	204 -4	169 -17	202 -8
2nd turn reflect on VE	174 -8	200 -8	180 -6	210 +0

Table 3: Limits of Single-Model Prompting and Reflection

Figure 3: M 3 -ACE Pipeline(right part) with Two Auxiliary Tools(left part).

Anchor Agent	CoT Infer	1st Regenerate with Summary	1st Refine		1st Reflect
Anchor Agent	CoT Infer	1st Regenerate with Summary	Select	Reject	Reject	All
Gemini 3 pro	85.0	88.1 3.1 $\uparrow$	92.7 2741	46.1 299	56.9 10.8 $\uparrow$	89.1 1.0 $\uparrow$
Gemini 2.5 pro	73.3	80.5 7.2 $\uparrow$	84.4 2812	32.2 228	41.6 9.4 $\uparrow$	81.2 0.7 $\uparrow$
GPT 5	72.0	81.3 9.3 $\uparrow$	85.5 2709	47.1 331	55.6 8.5 $\uparrow$	82.2 0.9 $\uparrow$
Claude-45	61.2	76.0 14.8 $\uparrow$	79.2 2250	66.7 790	68.7 1.0 $\uparrow$	76.5 0.5 $\uparrow$

Table 4: Results of the SOTA Ensemble. Subscripts without uparrows(

\uparrow

) denote the number of samples. The green numbers indicate the number of samples filtered as Select, while the red numbers indicate the number of samples filtered as Reject. The selected samples are merged into the final answer pool, whereas the rejected samples are regenerated and proceed to the next round. Subscripts with arrows represent the accuracy (ACC) improvement compared to the previous stage. The gray arrow subscripts indicate the ACC gain of the first-round “Regenerate with Summary” compared to CoT inference. The blue arrow subscripts denote the ACC improvement obtained after the 1st Reflect step on the rejected samples, relative to the answers from the first-round “Regenerate with Summary.” The purple arrow subscripts represent the overall ACC improvement on the entire dataset, from the first-round “Regenerate with Summary” to after the 1st Reflect stage.

Anchor Agent	CoT Infer	1st Regenerate with Summary
Gemini 2.5 pro	73.3	77.6
GPT 5	72.0	78.5
Claude-45	61.2	71.2

Table 5: Results of the sub-SOTA Ensemble.

Multimodal Reasoning

Multimodal reasoning studies how AI systems combine text with images, audio, video, and other signals to solve evidence-dependent tasks.

See all 508 papers

Jun 16, 2026

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

Chain-of-Thought (CoT) reasoning has extended from purely linguistic domains to multimodal scenarios; however, existing approaches often treat visual inputs as homogeneous or auxiliary signals, failing to capture the intricate and sample-specific dependencies between text and images in mathematical problem-solving. This gives rise to two core issues: first, the supervisory signals for visual content are generalized and coarse-grained, lacking adaptation to the actual necessity of visual information in each sample; second, training feedback becomes inaccurate when visual rewards are uniformly applied without distinguishing the complementary relationships among inputs. These limitations hinder models from achieving precise multimodal reasoning. In this work, we propose a framework for modeling fine-grained visual dependencies in mathematical reasoning. We first construct the MathVis-Fine dataset, augmenting fine-grained visual annotations with visual dependency ratings. Building upon this dataset, we introduce a two-stage progressive visual enhancement training paradigm that balances answer correctness rewards and visual grounding rewards according to the intrinsic visual dependency level of each sample, thereby mitigating reward bias and improving supervision accuracy. Extensive experiments demonstrate that the MathVis-Fine framework effectively enhances visual perception progressively based on visual dependency, offering a more precise training framework for multimodal mathematical reasoning. We will release the dataset upon acceptance.

Wanshi Xu, Haokun Zhao, Haidong Yuan +2

School of ECE, Peking University · College of Computer Science and Artificial Intelligence, Fudan University · School of Software and Microelectronics, Peking University +1

Multimodal Reasoning Multimodal Reward Modeling Multimodal CoT Reasoning Visually Grounded Reasoning Fine-Grained Visual Perception LLM Mathematical Reasoning

Ask PDF

Oct 18, 2026

MAMath-Bench: Process-Aware and Robust Evaluation of Multimodal Multi-Agent Mathematical Reasoning

Mathematical reasoning is a fundamental capability for educational applications of Large Language Models (LLMs), where reliable problem solving requires not only correct answers but also coherent and verifiable reasoning processes. Multi-agent systems (MASs) have shown growing potential in this domain by leveraging collaboration and verification among multiple LLM agents. However, existing evaluations primarily focus on the correctness of final answers in text-only settings, offering limited insight into multimodal reasoning capabilities, the reliability of intermediate reasoning steps, and robustness against reasoning errors. To address these limitations, we introduce MAMath-Bench, a process-aware benchmark designed to evaluate the multimodal mathematical reasoning capabilities of MASs. MAMath-Bench comprises 1,003 problems spanning 11 mathematical categories, multiple difficulty levels, and several modalities, including text, images, and audio. It features fine-grained milestone annotations to evaluate systems across three dimensions: problem-solving correctness, reasoning-process reliability, and robustness to adversarial reasoning-chain perturbations. Our evaluation yields three key findings. First, multi-agent collaboration does not consistently improve final-answer accuracy over strong single LLMs. Second, collaboration can substantially enhance the coverage of intermediate reasoning, but such gains may come at the cost of increased reasoning redundancy. Third, current systems remain vulnerable to adversarial reasoning errors, particularly those involving subtle structural perturbations. The dataset is available at https://github.com/mira-ai-lab/AgentMath.

Xinping Wang, Jie Ma, Rui Xing +6

School of Cyber Science and Engineering, Xi’an Jiaotong University · MOE KLINNS Lab, Xi’an Jiaotong University · School of Computer Science and Technology, Xi’an Jiaotong University +3

Mathematical Reasoning Benchmarks Language Model Reasoning Evaluation Multi-Agent Reasoning Multimodal Reasoning

Ask PDF

Sep 12, 2026

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

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.

Yuzhe Li, Hao Yan, Hao Wang +6

Huazhong University of Science and Technology · Huawei Inc.

VLM Reasoning Large Vision-Language Models Mathematical Reasoning Visually Grounded Reasoning RL for VLMs

Ask PDF

Abstract

Multimodal Reasoning

Explore similar work

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

MAMath-Bench: Process-Aware and Robust Evaluation of Multimodal Multi-Agent Mathematical Reasoning

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