Func-R1: Incentivizing Mathematical Function Reasoning in Multimodal Large Language Models
Mingze Yin, Xiaohan Wang, Dian Li, Haichao Yao, Yilin Zhao, Youjun Chen, Gang Liu, Jintai Chen, et al.
cs.CL
Quick verdict
Func-R1 decouples function-plot parsing from text reasoning via SFT+PATO RL, claiming open-source SOTA and +8.4% over GPT-5 on MathVerse; code unreleased.
Summary
Func-R1 targets a 'modality interference' problem the authors document in mathematical function reasoning: on MathVista subsets, GPT-5 reasons worse from function plots than from textual captions, and combining both yields the worst results (Figure 1). The proposed system decouples the pipeline into a visual math parser — a Qwen3-VL-4B/8B model fine-tuned with supervised learning on FuncCaption, a 3,519-problem corpus annotated by GPT-5/Gemini-3-Pro, arbitrated by other frontier models and validated by five mathematics PhDs — and a textual math reasoner (Qwen3-8B/32B) trained with reinforcement learning. The PATO algorithm modifies GRPO-style updates with a smooth, temperature-gated trust region and combines three rewards: LLM-judged visual-perception correction, knowledge-concept F1 alignment, and verifiable answer correctness. On out-of-distribution benchmarks (Table 1), the Pro variant reports 85.3 on MathVista and 82.6 on Math-Vision, best among tested models on 8 of 10 suites, with a claimed average 8.4% improvement over GPT-5 on MathVerse function tasks; the in-distribution average is 91.4% on the authors' own held-out split (Table 2). Transfer experiments (Table 3) report a 21% relative improvement over prior geometry SOTA on MathVista. Ablations credit each reward and show PATO beating GRPO/SAPO/DAPO (92.7 vs 87.4-90.2, Figure 5). Main caveats: code and data are withheld pending corporate review, key hyperparameters are unreported, and no significance testing accompanies the headline margins.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- MLLMs exhibit modality interference on function plots: GPT-5 reasons worse from diagrams than from textual captions, and joint image+caption input performs worst (Figure 1).
- A decoupled visual parser (SFT) plus textual reasoner (PATO RL) achieves the best open-source results across five math benchmarks, reportedly beating GPT-5 by an average 8.4% (relative) on MathVerse function tasks (Tables 1-2, Sec. 3.3).
- Each proposed reward component contributes measurably (Table 4: MathVista 65.7 to 73.1), and PATO outperforms GRPO, SAPO and DAPO under identical rewards and decoding (Figure 5: 92.7 vs 87.4-90.2).
- Function-centric training transfers across domains: GEO-R1/SYNSCI-R1 improve MathVista by about 21% relative over the prior geometry SOTA (Table 3: 75.9/81.8 vs 62.8).
Strengths
Concrete motivating diagnosis
Figure 1 shows GPT-5 on MathVista scoring worse with function diagrams than with captions and worst with both jointly, and Figure 6 attributes function-category errors mainly to visual misperception and knowledge deficit rather than computation deviation — a clear, testable phenomenon statement.
Reward design is ablated piecewise
Table 4 shows monotonic gains as rewards are added (MathVista 65.7 to 73.1; MathVerse 51.4 to 57.5), directly supporting the Rv/Rk/Rs decomposition of Eq. 5-8.
Controlled optimizer comparison
Broad evaluation plus transfer
Transparent annotation pipeline
Weaknesses
Artifacts currently unreproducible
Asymmetric in-distribution comparison
No variance or significance on the headline margins
Table ambiguity and unexplained config mismatch
Contamination risk unaddressed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper introduces FUNC-R1, a post-training recipe for multimodal mathematical function reasoning. The motivating observation (Sec. 1, Figure 1) is a modality-interference effect: on MathVista subsets, GPT-5 scores worse when given function diagrams than when given textual captions, and image+caption jointly performs worst. Claimed contributions are (i) identification of this interference, (ii) a 3,519-problem function-centric corpus (FuncCaption + FuncReasoning), (iii) PATO, a GRPO-derived policy optimizer with hybrid rewards, and (iv) open-source SOTA across five benchmarks, including an average relative +8.4% over GPT-5 on MathVerse function tasks (Sec. 3.3, Table 1) and a +21% relative gain over prior geometry SOTA on MathVista (Sec. 3.4, Table 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Func-R1?
Func-R1 is a multimodal LLM system for mathematical function reasoning built on Qwen3 models. A vision parser (Qwen3-VL, SFT-tuned) converts function plots into structured captions with described properties, and a text-only reasoner (Qwen3-8B/32B) solves the problem via reinforcement learning using the PATO algorithm and three reward signals (visual correction, knowledge alignment, verifiable answer).
Does Func-R1 beat GPT-5?
The authors report an average 8.4% improvement over GPT-5 on MathVerse function-oriented tasks and the best tested results in 8 of 10 benchmark suites (for example MathVista 85.3 vs 84.7). GPT-5 still leads on some suites, the margins are small without significance testing, and the comparisons come from the paper's own evaluation protocol.
Is the Func-R1 code or dataset publicly available?
Not yet. The paper points to a GitHub repository (zju-ai4s/Func-R1) that will be released only after a company disclosure review process; the 3,519-problem FuncCaption/FuncReasoning corpus is likewise withheld for now.
What is PATO in the Func-R1 paper?
Perception-Aligned Theoretic Optimization: a GRPO-style reinforcement learning method that replaces hard gradient clipping with a smooth, temperature-gated trust region (separate temperatures for positive and negative token updates) and combines rewards for visual-perception correction, knowledge-concept alignment (F1), and verifiable answer correctness.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Table 1 extraction shows 10 numeric values against a 12-label header, so per-column attribution of MathVerse vs We-Math subcategories is uncertain.
- Figures 1-6 are only partially recoverable from text; exact per-stage values in Figure 4 and per-algorithm values in Figure 5 may be mis-assigned.
- The GitHub repository's current status could not be checked (no network access during this review).
- Some row labels and model names appear garbled in extraction (e.g., variant naming in Table 1, 'Claude Opus 4.7' in Table 2).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.