Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Integrating Visual Interpretation and Linguistic Reasoning for Math Problem Solving

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Decoupling visual interpretation from linguistic reasoning outperforms end-to-end vision-language models on geometry problems.

desk verdict A clean three-stage recipe for decoupled vision-language reasoning with real gains on geometry, but the headline MathVerse numbers may be inflated by training-set overlap and the paper never checks. read the letter →

arxiv 2505.17609 v2 pith:CAH45CJZ submitted 2025-05-23 cs.AI

Zixian Guo , Ming Liu , Qilong Wang , Zhilong Ji , Jinfeng Bai , Lei Zhang , Wangmeng Zuo This is my paper
classification cs.AI
keywords vision-languagemodelsgeometricmathproblemsdecoupledreasoningimageinterpretationoutcome-rewardedtuningGRPOVerseLVLM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the hard part of visual math problems is not reasoning but getting accurate information out of the image, and that this should be handled by a separate specialist rather than by an end-to-end multimodal model. It proposes a decoupled framework: a vision-language model (LVLM) converts the diagram into a textual description, then a text-only LLM solves the problem from that description and the question. The two modules are first fine-tuned separately, then jointly tuned with an outcome-rewarded reinforcement step in which the correctness of the final answer is the only reward. On the MathVerse geometry benchmark, the resulting 7B+7B system scores 47.2% on the vision-dominant split and 43.8% on the vision-only split, compared with 38.5% and 28.5% for the R-CoT-7B baseline. If this holds, it suggests a cheaper path to multimodal reasoning: keep improving text reasoners and pair them with specialized image interpreters instead of training ever-larger end-to-end vision-language models.

What carries the argument

The load-bearing mechanism is an explicit text description inserted between vision and reasoning. The LVLM (Qwen2VL-7B) is trained to output a compact natural-language account of geometric primitives and relations; the LLM (Qwen2.5Math-7B) never sees the image. The three-stage procedure is the machinery that makes this work: (1) supervised fine-tuning of both models on GeoMM-derived data, (2) GRPO-style reinforcement of the LVLM using the final-answer correctness of the frozen LLM as reward, and (3) GRPO-style reinforcement of the LLM using the same outcome reward while the LVLM is frozen. The reward signal flows from a rule-based comparison of the final answer with ground truth, so the two modules are optimized as one chained policy.

What would settle it

Run an image and question overlap audit between the GeoMM/Geo170k training data and MathVerse testmini, then re-evaluate with any overlapping samples removed; if the Vision Dominant margin over R-CoT-7B disappears, the decoupling claim would be refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a division of labor—an LVLM that describes, an LLM that reasons—can outperform the usual end-to-end LVLM on visually intensive geometry problems. The authors construct an image-description dataset from GeoMM, refine the rule-generated descriptions with GPT-4o, embed question text into images for training, and fine-tune Qwen2VL-7B to read and describe diagrams and Qwen2.5Math-7B to reason from the resulting text. They then run two reinforcement stages: first the LVLM is rewarded when the frozen LLM gives the correct answer from the LVLM's description; then the LLM is rewarded with the LVLM frozen. In Table 2, the trained 7B+7B system reaches 47.2% on MathVerse Vision Dominant and 43.8% on Vision Only, versus 38.5% and 28.5% for R-CoT-7B, and comes close to a much larger Qwen2VL-72B + DeepSeek-V3 pipeline. The same recipe applied to LLaVA-o1 data also beats LLaVA-o1 itself, which the authors read as evidence that separating interpretation from reasoning, rather than training one model to do both, carries the gain.

Load-bearing premise

The reported numbers assume the GeoMM and Geo170k training samples do not overlap with the MathVerse testmini set; the paper does not describe a deduplication or contamination check.

Editorial extensions

If this is right

  • A 7B+7B decoupled system can reach or approach the performance of pipelines dozens of times larger on geometry benchmarks, so model size is not the only route to visual math ability.
  • Because the geometric image is reduced to text, swapping in a newer or stronger text-only LLM is a direct upgrade path: the visual interpreter stays fixed and only the reasoner is retrained.
  • The largest gains appear on the MathVerse splits that force the model to rely on the image, matching the claim that the bottleneck in end-to-end LVLMs is visual extraction, not reasoning.
  • The same decoupled recipe applied to LLaVA-o1's training data outperforms LLaVA-o1 itself, suggesting the advantage transfers to at least one other dataset's question style.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An immediate testable extension is to apply the same two-module recipe to other geometry benchmarks and to non-geometry visual reasoning, since the paper only evaluates on math and general VQA sets.
  • Because the reward is a single final-answer match, the LVLM could in principle learn descriptions that exploit reasoning shortcuts of the LLM without faithfully representing the diagram; measuring description fidelity against ground-truth caption data would separate these two effects.
  • The authors do not report a contamination check between the GeoMM/Geo170k training sets and MathVerse testmini; a duplicate-image audit would establish whether the headline margins hold when overlap is removed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a decoupled vision-language reasoning framework (DVLR) that separates visual interpretation from linguistic reasoning: a Qwen2VL-7B LVLM converts geometry diagram images into structured textual descriptions, and a Qwen2.5Math-7B LLM produces step-by-step solutions from those descriptions and the original question. Training has three stages: supervised fine-tuning of both modules on image-description and reasoning data derived from GeoMM; outcome-rewarded GRPO fine-tuning of the LVLM using the answer correctness of the frozen LLM; and outcome-rewarded fine-tuning of the LLM with the LVLM frozen. The authors evaluate on MathVerse testmini, MathVista, MMStar, and MM-Vet, reporting the largest gains on the vision-dominant and vision-only subsets of MathVerse, where the 7B+7B system reaches 47.2% and 43.8% versus 38.5% and 28.5% for R-CoT-7B. Ablations cover description preprocessing, decoupling versus a trained LVLM-only system, and the contribution of each training stage; code is released.

Significance. If the empirical results hold, this is a useful and cost-effective contribution to vision-language math reasoning. The paper demonstrates that a pipelined pair of existing 7B models with targeted outcome-rewarded tuning can outperform end-to-end LVLMs on vision-heavy geometry benchmarks, and the released code plus the stage-wise ablations make the approach easy to reproduce and extend. The contribution is empirical; its central claim rests entirely on the MathVerse comparison, so the missing contamination check and the absent no-training decoupled baseline in the same model family are the decisive gaps. With those gaps filled, the paper would be a solid benchmark contribution rather than an incremental system paper.

major comments (4)
  1. [Section 4.1 and Table 2] The training data (GeoMM, 33,344 images/86,857 QA pairs; Geo170k, 8,063 images/117,205 QA pairs) and the MathVerse testmini evaluation are never checked for overlap. Both Geo170k and MathVerse are built in part from public geometry QA sources such as GeoQA and Geometry3K, so it is plausible that some MathVerse testmini samples appear in the SFT or RL data. Because Stages 2 and 3 (Sections 3.3 and 3.4) use a rule-based reward that compares the extracted final answer with the ground-truth answer, any duplicated sample would be rewarded for reproducing the benchmark answer, inflating the Vision Dominant (47.2 vs. 38.5) and Vision Only (43.8 vs. 28.5) gains over R-CoT-7B. The authors should perform image-level and text-level deduplication against MathVerse and re-report results on the non-overlapping subset, or quantify the overlap and show that the conclusions are unchanged.
  2. [Section 4.3 and Table 2] The comparison lacks an off-the-shelf decoupled baseline at the same model scale. The paper includes Qwen2VL-72B & Deepseek-V3(671B) and GPT-4o & OpenAI-o1mini as decoupled pipelines, but these are much larger or closed-source, and Table 4's LVLM-only ablation is trained, not an untrained captioner. To establish that the three-stage joint tuning is what produces the gains, the authors should evaluate a zero-shot (no fine-tuning) pipeline consisting of Qwen2VL-7B producing image descriptions followed by Qwen2.5Math-7B reasoning. Without this baseline, the contribution of the proposed training is confounded with the benefit of the decoupled architecture itself.
  3. [Sections 4.2-4.4] All comparisons report a single accuracy number with no variance estimate, confidence interval, or significance test. Since the RL stages are stochastic and MathVerse testmini is a finite benchmark, the differences that motivate the headline claim, for example 47.2 vs. 44.2 for Ours vs. Qwen2VL-72B & Deepseek-V3 on Vision Dominant, are not shown to be distinguishable from sampling noise. The authors should report at least three seeds or a bootstrap confidence interval for the main results.
  4. [Section 3.3, Eq. (1)-(3)] The KL regularizer in Eq. (3) is written as pi_ref/pi_theta - log(pi_ref/pi_theta) - 1, which is not the standard form of D_KL(pi_theta||pi_ref) used in most GRPO implementations, and the text never specifies the reference policy used in Stages 2 and 3 (e.g., whether pi_ref is the SFT checkpoint or a previous RL iteration). Without this information, the optimization procedure cannot be reproduced exactly, which matters because the paper's central claim is about the effect of joint tuning.
minor comments (5)
  1. [Sections 3.2 and 4.4] The text refers to 'Section 2.2' when describing augmented GeoMM images and to 'Section 2.4' when describing the RL strategy; these should be Sections 3.2 and 3.4, respectively.
  2. [References] References [17] and [18] both cite the same paper, 'Visual instruction tuning'; this duplication should be corrected, and the intended second reference (likely LLaVA-1.5) should be supplied.
  3. [Section 4.3] The parenthetical '(e.g., "Vision Only" -23.8)' is undefined; it presumably means the drop from Text Dominant to Vision Only for R-CoT-7B, but this should be stated explicitly.
  4. [Figure 2 caption] The caption for Figure 2 is difficult to parse ('Image Image VLM LM VLM LM...') and should be rewritten as a clear description of the three-stage pipeline.
  5. [Table 2] The table would be more informative if it marked which compared systems are decoupled (LVLM + LLM) versus end-to-end LVLMs, and if it reported the number of test samples in each MathVerse subset, since percentage differences on small subsets can be noisy.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity was found: the method is empirical, the benchmarks are external, and no claim in the paper reduces to its inputs by construction.

full rationale

The paper makes no derivation that reduces to its inputs. Its contribution is an empirical training pipeline: supervised fine-tuning on GeoMM and Geo170k, followed by GRPO-based reinforcement learning with a rule-based outcome reward (Eqs. 1-3), and evaluation on external benchmarks (MMStar, MMVet, MathVista, MathVerse). The reward is answer correctness on the training distribution, which is the standard objective of supervised or reinforcement fine-tuning, not a circular use of the test quantity. The MathVerse scores in Table 2 are held-out measurements of the trained pipeline, not quantities constructed from the training equations. Ablation studies in Tables 3-5 isolate the contributions of description preprocessing, decoupling, and the two reinforcement stages, and the comparisons in Tables 1-2 use external baselines. The cited datasets and benchmarks are prior external work, not self-citations; no uniqueness theorem or ansatz is imported from the authors' own prior work, and no architectural choice is justified solely by a citation. The only substantive concern is possible train/test overlap between Geo170k or GeoMM and the MathVerse testmini set, since all draw on public geometry QA sources; if specific testmini samples appeared in training, the reported gains would be inflated by memorization. However, the paper provides no evidence that any particular MathVerse testmini sample is in the training data, and a potential contamination risk is a data-hygiene or correctness issue, not a circular step under the required standard of exhibiting a specific reduction. Therefore, no significant circularity is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new theoretical entities. It relies on standard assumptions about the quality of synthetic geometry datasets, the adequacy of binary reward for RL, and the reasoning capability of the base LLM. The most fragile assumption is the absence of data leakage between training and evaluation benchmarks.

free parameters (4)
  • GRPO clip epsilon = 0.2
    Chosen by hand for the reinforcement learning stage; standard value from existing GRPO implementations, not tuned on the benchmark.
  • KL coefficient beta = 0.01
    Chosen by hand to control deviation from the reference policy; not tuned on the benchmark.
  • GRPO group size G = 5
    Chosen by hand for sampling descriptions/responses per question; affects variance of advantage estimates.
  • RL learning rate = 1e-6
    Chosen by hand; standard small learning rate for RL fine-tuning of large models.
assumptions (4)
  • domain assumption GeoMM's rule-generated image descriptions are accurate and contain all necessary visual information for solving the geometry problems.
    The method relies on these descriptions as supervision for the VLM's interpretation ability. Any systematic errors in the labels would propagate to the trained model.
  • domain assumption The outcome reward (correct/incorrect final answer from the LLM) is a sufficient signal to improve the VLM's description generation for the LLM.
    Stage 2 optimizes the VLM solely from this binary reward; if the reward is too sparse or noisy, the RL update may not improve descriptions.
  • domain assumption Qwen2.5Math-7B is capable enough to reason from textual descriptions of diagrams.
    The whole architecture depends on the LLM being able to solve geometry from text alone, which is plausible but not demonstrated in isolation.
  • domain assumption MathVerse testmini does not overlap with GeoMM/Geo170k training data.
    The paper provides no contamination analysis; if overlap exists, the reported accuracy is inflated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating Visual Interpretation and Linguistic Reasoning for Math Problem Solving." pith.science (2026). https://pith.science/paper/CAH45CJZ

@misc{pith2026250517609,
  author       = {Pith},
  title        = {Pith review of: Integrating Visual Interpretation and Linguistic Reasoning for Math Problem Solving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CAH45CJZ}},
  note         = {Machine review of arXiv:2505.17609}
}
read the original abstract

Current large vision-language models (LVLMs) typically employ a connector module to link visual features with text embeddings of large language models (LLMs) and use end-to-end training to achieve multi-modal understanding in a unified process. Effective alignment needs high-quality pre-training data and a carefully designed training process. Current LVLMs face challenges when addressing complex vision-language reasoning tasks, with their reasoning capabilities notably lagging behind those of LLMs. This paper proposes a paradigm shift: instead of training end-to-end vision-language reasoning models, we advocate for developing a decoupled reasoning framework based on existing visual interpretation specialists and text-based reasoning LLMs. Our approach leverages (1) a dedicated vision-language model to transform the visual content of images into textual descriptions and (2) an LLM to perform reasoning according to the visual-derived text and the original question. This method presents a cost-efficient solution for multi-modal model development by optimizing existing models to work collaboratively, avoiding end-to-end development of vision-language models from scratch. By transforming images into language model-compatible text representations, it facilitates future low-cost and flexible upgrades to upcoming powerful LLMs. We introduce an outcome-rewarded joint-tuning strategy to optimize the cooperation between the visual interpretation and linguistic reasoning model. Evaluation results on vision-language benchmarks demonstrate that the decoupled reasoning framework outperforms recent LVLMs. Our approach yields particularly significant performance gains on visually intensive geometric mathematics problems. The code is available: https://github.com/guozix/DVLR.

Figures

Figures reproduced from arXiv: 2505.17609 by the authors.

Figure 1
Figure 1. Comparison of the solving results of a geometry math problem by directly answering with a VL model or first interpreting the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The three-stage model training pipeline of our method. The first supervised fine-tuning stage establishes a foundation of both [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Demonstration of the geometry image description data. The original image description from the GeoMM dataset is accurate but [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of different versions of the same problem [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The output examples of decoupled vision-language reasoning method. The left example shows that reinforced finetuning can [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Image interpretation result of the LVLM if preprocessing [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CPPO: Contrastive Perception Policy Optimization for VLM Agents

    cs.CV 2026-01 conditional novelty 6.0 of 10

    CPPO improves VLM agents by adding a contrastive perception loss to GRPO, applied only to entropy-selected image-dependent tokens, beating prior RL finetuning methods on visual reasoning benchmarks.

Reference graph

Works this paper leans on

46 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men- sch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716–23736,

  2. [2]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3

    arXiv 2023

  3. [3]

    Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330,

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330,

    arXiv

  4. [4]

    Comt: A novel benchmark for chain of multi-modal thought on large vision- language models, 2024

    Zihui Cheng, Qiguang Chen, Jin Zhang, Hao Fei, Xiaocheng Feng, Wanxiang Che, Min Li, and Libo Qin. Comt: A novel benchmark for chain of multi-modal thought on large vision- language models, 2024. 1

    work page 2024

  5. [5]

    R-cot: Reverse chain-of-thought problem generation for geometric reasoning in large multimodal models, 2024

    Linger Deng, Yuliang Liu, Bohan Li, Dongliang Luo, Liang Wu, Chengquan Zhang, Pengyuan Lyu, Ziyang Zhang, Gang Zhang, Errui Ding, Yingying Zhu, and Xiang Bai. R-cot: Reverse chain-of-thought problem generation for geometric reasoning in large multimodal models, 2024. 3, 4, 5, 6, 7

    work page 2024

  6. [6]

    G-llava: Solving geomet- ric problem with multi-modal large language model, 2023

    Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, and Lingpeng Kong. G-llava: Solving geomet- ric problem with multi-modal large language model, 2023. 3, 4, 5, 6

    work page 2023

  7. [7]

    Interleaved-modal chain-of-thought, 2024

    Jun Gao, Yongqi Li, Ziqiang Cao, and Wenjie Li. Interleaved-modal chain-of-thought, 2024. 1

    work page 2024

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1

    arXiv 2025

Show all 46 references
  1. [9]

    Infimm-webmath-40b: Ad- vancing multimodal pre-training for enhanced mathematical reasoning, 2024

    Xiaotian Han, Yiren Jian, Xuefeng Hu, Haogeng Liu, Yiqi Wang, Qihang Fan, Yuang Ai, Huaibo Huang, Ran He, Zhen- heng Yang, and Quanzeng You. Infimm-webmath-40b: Ad- vancing multimodal pre-training for enhanced mathematical reasoning, 2024. 3

    2024

  2. [10]

    Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark, 2025

    Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark, 2025. 1

    2025

  3. [11]

    V- star: Training verifiers for self-taught reasoners, 2024

    Arian Hosseini, Xingdi Yuan, Nikolay Malkin, Aaron Courville, Alessandro Sordoni, and Rishabh Agarwal. V- star: Training verifiers for self-taught reasoners, 2024. 3

    2024

  4. [12]

    Why vision language models struggle with visual arithmetic? to- wards enhanced chart and geometry understanding, 2025

    Kung-Hsiang Huang, Can Qin, Haoyi Qiu, Philippe Laban, Shafiq Joty, Caiming Xiong, and Chien-Sheng Wu. Why vision language models struggle with visual arithmetic? to- wards enhanced chart and geometry understanding, 2025. 1

    2025

  5. [13]

    Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M

    Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D. Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M. Zhang, Kay McK- inney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, and Aleksandra Faust. ...

    2024

  6. [14]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 3

    2022

  7. [15]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In In- ternational conference on machine learning , pages 19730– 19742. PMLR, 2023. 1, 3

    2023

  8. [16]

    Let’s verify step by step,

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Ed- wards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step,

  9. [17]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892–34916, 2023. 3

    2023

  10. [18]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892–34916, 2023. 1, 3

    2023

  11. [19]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024. 3

    2024

  12. [20]

    Mathvista: Evaluating mathemat- ical reasoning of foundation models in visual contexts, 2024

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathemat- ical reasoning of foundation models in visual contexts, 2024. 6

    2024

  13. [21]

    Reft: Reasoning with reinforced fine-tuning, 2024

    Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning, 2024. 1, 3, 5

    2024

  14. [22]

    Exploring the limit of outcome reward for learning mathematical reasoning, 2025

    Chengqi Lyu, Songyang Gao, Yuzhe Gu, Wenwei Zhang, Jianfei Gao, Kuikun Liu, Ziyi Wang, Shuaibin Li, Qian Zhao, Haian Huang, Weihan Cao, Jiangning Liu, Hong- wei Liu, Junnan Liu, Songyang Zhang, Dahua Lin, and Kai Chen. Exploring the limit of outcome reward for learning mathema...

    2025

  15. [23]

    S1: Simple test-time scaling, 2025

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Cand `es, and Tatsunori Hashimoto. S1: Simple test-time scaling, 2025. 3

    2025

  16. [24]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anad- kat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Bal- com, Paul Baltescu, Haiming Bao, Mohammad Bavarian, J...

    2024

  17. [25]

    Multimath: Bridging visual and math- ematical reasoning for large language models, 2024

    Shuai Peng, Di Fu, Liangcai Gao, Xiuqin Zhong, Hongguang Fu, and Zhi Tang. Multimath: Bridging visual and math- ematical reasoning for large language models, 2024. 3, 4, 6

    2024

  18. [26]

    Mutual reasoning makes smaller llms stronger problem-solvers, 2024

    Zhenting Qi, Mingyuan Ma, Jiahang Xu, Li Lyna Zhang, Fan Yang, and Mao Yang. Mutual reasoning makes smaller llms stronger problem-solvers, 2024. 3

    2024

  19. [27]

    O1 replication journey: A strategic progress report – part 1, 2024

    Yiwei Qin, Xuefeng Li, Haoyang Zou, Yixiu Liu, Shijie Xia, Zhen Huang, Yixin Ye, Weizhe Yuan, Hector Liu, Yuanzhi Li, and Pengfei Liu. O1 replication journey: A strategic progress report – part 1, 2024. 3

    2024

  20. [28]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728–53741, 2023. 1

    2023

  21. [29]

    Deepseekmath: Pushing the limits of mathe- matical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathe- matical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 1, 3, 4

    2024 arXiv

  22. [30]

    Math- llava: Bootstrapping mathematical reasoning for multimodal large language models, 2024

    Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math- llava: Bootstrapping mathematical reasoning for multimodal large language models, 2024. 3, 4, 6

    2024

  23. [31]

    Scaling llm test-time compute optimally can be more effec- tive than scaling model parameters, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effec- tive than scaling model parameters, 2024. 3

    2024

  24. [32]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 5

    2024 arXiv

  25. [33]

    Peiyi Wang, Lei Li, Zhihong Shao, R. X. Xu, Damai Dai, Yifei Li, Deli Chen, Y . Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human anno- tations, 2024. 3

    2024

  26. [34]

    Slow perception: Let’s perceive geometric figures step-by-step,

    Haoran Wei, Youyang Yin, Yumeng Li, Jia Wang, Liang Zhao, Jianjian Sun, Zheng Ge, and Xiangyu Zhang. Slow perception: Let’s perceive geometric figures step-by-step,

  27. [35]

    Gee, and Yixin Nie

    Yifan Wu, Pengchuan Zhang, Wenhan Xiong, Barlas Oguz, James C. Gee, and Yixin Nie. The role of chain-of-thought in complex vision-language reasoning task, 2023. 1

    2023

  28. [36]

    Atomthink: A slow thinking framework for multi- modal mathematical reasoning, 2024

    Kun Xiang, Zhili Liu, Zihao Jiang, Yunshuang Nie, Run- hui Huang, Haoxiang Fan, Hanhui Li, Weiran Huang, Yihan Zeng, Jianhua Han, Lanqing Hong, Hang Xu, and Xiaodan Liang. Atomthink: A slow thinking framework for multi- modal mathematical reasoning, 2024. 1

    2024

  29. [37]

    Llava-o1: Let vision language models reason step- by-step, 2024

    Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-o1: Let vision language models reason step- by-step, 2024. 1, 3, 6

    2024

  30. [38]

    Mulberry: Em- powering mllm with o1-like reasoning and reflection via col- lective monte carlo tree search, 2024

    Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, and Dacheng Tao. Mulberry: Em- powering mllm with o1-like reasoning and reflection via col- lective monte carlo tree search, 2024. 1, 3

    2024

  31. [39]

    Metamath: Bootstrap your own mathe- matical questions for large language models

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathe- matical questions for large language models. arXiv preprint arXiv:2309.12284, 2023. 3

    2023 arXiv

  32. [40]

    Mm-vet: Evaluating large multimodal models for integrated capabilities

    Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 6

    2023 arXiv

  33. [41]

    Llama-berry: Pair- wise optimization for o1-like olympiad-level mathematical reasoning, 2024

    Di Zhang, Jianbo Wu, Jingdi Lei, Tong Che, Jiatong Li, Tong Xie, Xiaoshui Huang, Shufei Zhang, Marco Pavone, Yuqiang Li, Wanli Ouyang, and Dongzhan Zhou. Llama-berry: Pair- wise optimization for o1-like olympiad-level mathematical reasoning, 2024. 3

    2024

  34. [42]

    Euclid: Supercharging multimodal llms with synthetic high-fidelity visual descriptions, 2024

    Jiarui Zhang, Ollie Liu, Tianyu Yu, Jinyi Hu, and Willie Neiswanger. Euclid: Supercharging multimodal llms with synthetic high-fidelity visual descriptions, 2024. 1

    2024

  35. [43]

    Mathverse: Does your multi-modal llm truly see the diagrams in visual math prob- lems?, 2024

    Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, and Hongsheng Li. Mathverse: Does your multi-modal llm truly see the diagrams in visual math prob- lems?, 2024. 6

    2024

  36. [44]

    Improve vision language model chain-of- thought reasoning

    Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of- thought reasoning. arXiv preprint arXiv:2410.16198, 2024. 1

    2024 arXiv

  37. [45]

    How far are we from intelligent visual deductive reasoning?, 2024

    Yizhe Zhang, He Bai, Ruixiang Zhang, Jiatao Gu, Shuangfei Zhai, Josh Susskind, and Navdeep Jaitly. How far are we from intelligent visual deductive reasoning?, 2024. 1

    2024

  38. [46]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.