arxiv: 2604.18584 · v1 · submitted 2026-04-20 · 💻 cs.AI · cs.DL· cs.IR· cs.LG

Recognition: unknown

MathNet: a Global Multimodal Benchmark for Mathematical Reasoning and Retrieval

Shaden Alshammari , Kevin Wen , Abrar Zainal , Mark Hamilton , Navid Safaei , Sultan Albarakati , William T. Freeman , Antonio Torralba

Authors on Pith no claims yet

Pith reviewed 2026-05-10 04:06 UTC · model grok-4.3

classification 💻 cs.AI cs.DLcs.IRcs.LG

keywords mathematical reasoningOlympiad problemsmultimodal datasetproblem retrievalretrieval-augmented generationmultilingual benchmarkAI evaluation

0 comments

The pith

MathNet dataset shows even top AI models struggle with Olympiad math problems and retrieval.

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

The paper introduces MathNet, a large-scale dataset of over 30,000 Olympiad-level math problems from 47 countries and 17 languages, along with a new benchmark for mathematical retrieval. It evaluates generative models on problem solving and embedding models on retrieving equivalent problems. Results indicate that leading models like Gemini achieve 78.4% and GPT 69.3% on solving tasks, with retrieval models performing poorly, though retrieval-augmented generation can improve results by up to 12%. This benchmark matters as it tests AI on diverse, high-level math across cultures and provides a standard for future improvements in reasoning and retrieval capabilities.

Core claim

MathNet is the largest high-quality Olympiad dataset with the first benchmark for mathematical problem retrieval using human-curated equivalent and similar problem pairs. State-of-the-art reasoning models remain challenged on the tasks, while retrieval-augmented generation yields performance gains of up to 12% when retrieval quality is high.

What carries the argument

The MathNet dataset and its retrieval benchmark of mathematically equivalent and structurally similar problem pairs curated by human experts.

If this is right

State-of-the-art models need significant improvements to reliably solve Olympiad-level problems.
Retrieval quality is critical for the success of retrieval-augmented approaches in mathematical tasks.
The multilingual aspect allows evaluation of models across different languages and cultural contexts in math.
Embedding models require better methods to capture mathematical equivalence beyond surface similarity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

MathNet could serve as a training resource to develop specialized mathematical AI systems.
Similar benchmarks might be developed for other domains like physics or programming to test retrieval-augmented reasoning.
Automated methods for generating equivalent problems could help scale such datasets further without heavy human curation.

Load-bearing premise

Human experts accurately and consistently identify pairs of mathematically equivalent and structurally similar problems without selection bias or judgment errors.

What would settle it

A new model achieving over 95% accuracy on the problem solving benchmark or perfect retrieval of all equivalent problems would indicate the challenges are overstated.

Figures

Figures reproduced from arXiv: 2604.18584 by Abrar Zainal, Antonio Torralba, Kevin Wen, Mark Hamilton, Navid Safaei, Shaden Alshammari, Sultan Albarakati, William T. Freeman.

**Figure 1.** Figure 1: Overview of MATHNET. MATHNET contains 30K+ Olympiad-level problems across 47 countries, 17 languages, and 143 competitions over 40 years with expert-authored solutions. We evaluate several leading models on problem solving and math-aware retrieval. ABSTRACT Mathematical problem solving remains a challenging test of reasoning for large language and multimodal models, yet existing benchmarks are limited in s… view at source ↗

**Figure 2.** Figure 2: Overview of MathNet-Solve. The dataset spans national, regional, TST, and international competitions, with varying solution lengths. It has grown since the early 2000s and includes textual and diagram-based problems with broad multilingual and topical coverage. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of MATHNET problem–solution extraction pipeline. The curation pipeline consists of three stages: (1) document ingestion and problem segmentation, (2) problem and solution extraction with format normalization, and (3) multi-stage extraction verification. 3.2 DATA COLLECTION, EXTRACTION AND ANNOTATION Data sources. Each year, participating IMO countries contribute original problems for use in their … view at source ↗

**Figure 4.** Figure 4: MathNet is a collection of official Olympiad documents sourced directly from national problem booklets. This example shows a BMO 2023 problem that appears in both MathNet and Omni-MATH Gao et al. (2024a) While Omni-MATH relies on the AoPS discussion shown on the left, MathNet provides the official problem and solution on the right. Problem Solving Accuracy (%, ↑) on MathNet-Solve-Test subsets Model Full Se… view at source ↗

**Figure 5.** Figure 5: Examples of scanned pages from national mathematics Olympiad booklets across different [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗

**Figure 6.** Figure 6: Cosine similarity distributions for equivalent (green) and near-miss/hard negatives (orange) [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗

read the original abstract

Mathematical problem solving remains a challenging test of reasoning for large language and multimodal models, yet existing benchmarks are limited in size, language coverage, and task diversity. We introduce MathNet, a high-quality, large-scale, multimodal, and multilingual dataset of Olympiad-level math problems together with a benchmark for evaluating mathematical reasoning in generative models and mathematical retrieval in embedding-based systems. MathNet spans 47 countries, 17 languages, and two decades of competitions, comprising 30,676 expert-authored problems with solutions across diverse domains. In addition to the core dataset, we construct a retrieval benchmark consisting of mathematically equivalent and structurally similar problem pairs curated by human experts. MathNet supports three tasks: (i) Problem Solving, (ii) Math-Aware Retrieval, and (iii) Retrieval-Augmented Problem Solving. Experimental results show that even state-of-the-art reasoning models (78.4% for Gemini-3.1-Pro and 69.3% for GPT-5) remain challenged, while embedding models struggle to retrieve equivalent problems. We further show that retrieval-augmented generation performance is highly sensitive to retrieval quality; for example, DeepSeek-V3.2-Speciale achieves gains of up to 12%, obtaining the highest scores on the benchmark. MathNet provides the largest high-quality Olympiad dataset together with the first benchmark for evaluating mathematical problem retrieval, and we publicly release both the dataset and benchmark at https://mathnet.mit.edu.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MathNet delivers a large new multilingual Olympiad dataset and retrieval benchmark, but the human-curated equivalence pairs lack the validation details needed to trust the retrieval and RAG results.

read the letter

MathNet's core value is the release of 30,676 Olympiad problems spanning 17 languages and 47 countries, plus a retrieval benchmark built from human-curated pairs of mathematically equivalent and structurally similar problems. This extends beyond English-only sets like MATH in both scale and language coverage, and the public data release at mathnet.mit.edu is the part that actually moves the field forward. The problem-solving numbers are straightforward: Gemini-3.1-Pro reaches 78.4% and GPT-5 reaches 69.3%, which lines up with what we already know about current model limits on hard math. The RAG experiments show retrieval quality matters, with gains up to 12% for some models when good retrieval is added. That part is useful to see quantified. The retrieval benchmark itself is the novel piece, and embedding models do struggle on it as claimed. The soft spot is exactly where the stress-test flagged: the equivalence and similarity pairs are described as human-expert curated, but the paper gives no criteria for what counts as equivalent, no count of annotators per pair, and no inter-annotator agreement numbers. Without those, the measured retrieval difficulty and the reported RAG sensitivity rest on unverified labels, which weakens the retrieval claims more than the solving ones. The dataset release still stands on its own. This is the kind of paper that belongs in a reading group for anyone working on math reasoning or math-aware retrieval, because the data itself is new and large enough to be worth testing against. It deserves peer review so reviewers can ask for the missing curation details and check whether the benchmark holds up once those are supplied.

Referee Report

2 major / 1 minor

Summary. The paper introduces MathNet, a large-scale, multimodal, multilingual dataset comprising 30,676 expert-authored Olympiad-level math problems from 47 countries and 17 languages over two decades. It establishes benchmarks for three tasks: problem solving, math-aware retrieval using human-curated pairs of mathematically equivalent and structurally similar problems, and retrieval-augmented problem solving. The results demonstrate that state-of-the-art models achieve at most 78.4% on problem solving (Gemini-3.1-Pro) and 69.3% (GPT-5), embedding models struggle with retrieval, and retrieval-augmented generation can provide gains of up to 12%, with the dataset and benchmark publicly released.

Significance. Should the human curation of the retrieval pairs prove reliable upon validation, this work would offer significant value by providing the largest high-quality Olympiad math dataset to date and the first benchmark specifically for mathematical problem retrieval. The findings on model challenges and the benefits of RAG in this domain could inform advancements in mathematical reasoning systems. The public release supports further research and reproducibility in the field.

major comments (2)

[Retrieval benchmark construction] The manuscript describes the retrieval benchmark as consisting of 'mathematically equivalent and structurally similar problem pairs curated by human experts' but does not specify the criteria for equivalence, the number of experts per pair, or inter-annotator agreement metrics. This information is essential to substantiate the claims regarding the difficulty for embedding models and the sensitivity of RAG performance to retrieval quality, as unvalidated labels may introduce bias or noise affecting the reported results.
[Experimental results on RAG] The performance gains from retrieval-augmented generation, such as the up to 12% improvement for DeepSeek-V3.2-Speciale, are reported without accompanying statistical significance tests, standard errors, or details on evaluation variance across multiple runs. This makes it difficult to determine whether the gains are robust or could be due to chance, particularly given the central role of RAG in the benchmark.

minor comments (1)

[Abstract] The mention of 'GPT-5' should be clarified, as it may refer to a specific model version or be a typographical reference to an existing model like GPT-4o.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for their detailed and constructive review of our manuscript. Their comments have helped us identify areas where the paper can be strengthened, particularly regarding the transparency of the retrieval benchmark construction and the statistical rigor of the RAG experiments. We address each major comment below and commit to making the necessary revisions.

read point-by-point responses

Referee: [Retrieval benchmark construction] The manuscript describes the retrieval benchmark as consisting of 'mathematically equivalent and structurally similar problem pairs curated by human experts' but does not specify the criteria for equivalence, the number of experts per pair, or inter-annotator agreement metrics. This information is essential to substantiate the claims regarding the difficulty for embedding models and the sensitivity of RAG performance to retrieval quality, as unvalidated labels may introduce bias or noise affecting the reported results.

Authors: We thank the referee for pointing out this gap in our description. The manuscript indeed provides only a brief mention of the human curation without elaborating on the process. We will revise the relevant section to include a detailed account of the curation protocol. This will specify the criteria for mathematical equivalence and structural similarity, the number of experts per pair, and the inter-annotator agreement metrics from our curation process. These additions will provide the necessary validation for the benchmark's quality and support our conclusions on model performance and RAG sensitivity. revision: yes
Referee: [Experimental results on RAG] The performance gains from retrieval-augmented generation, such as the up to 12% improvement for DeepSeek-V3.2-Speciale, are reported without accompanying statistical significance tests, standard errors, or details on evaluation variance across multiple runs. This makes it difficult to determine whether the gains are robust or could be due to chance, particularly given the central role of RAG in the benchmark.

Authors: We agree that reporting statistical measures is crucial for establishing the reliability of the RAG gains. The current results are based on single-run evaluations, which limits the assessment of variance. In the revised manuscript, we will include standard errors and results from statistical significance tests computed over multiple evaluation runs to demonstrate the robustness of the gains, ensuring the improvements are not due to chance. revision: yes

Circularity Check

0 steps flagged

No circularity: benchmark and results are direct empirical evaluations on newly collected data.

full rationale

The paper introduces a new dataset of 30,676 Olympiad problems and a retrieval benchmark built from human-curated equivalent/similar pairs, then reports direct model performance numbers (e.g., Gemini 78.4%, RAG gains up to 12%). No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. All claims rest on the external collection and evaluation process rather than any internal reduction to the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a dataset curation and benchmarking paper; it introduces no free parameters, mathematical axioms, or invented entities, relying instead on aggregation of existing competition problems.

pith-pipeline@v0.9.0 · 5599 in / 1026 out tokens · 29388 ms · 2026-05-10T04:06:31.813723+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

64 extracted references · 20 canonical work pages · 9 internal anchors

[1]

Omni-MATH: A Universal Olympiad Level Mathematic Benchmark For Large Language Models

Omni-MATH: A Universal Olympiad Level Mathematic Benchmark For Large Language Models , author=. arXiv preprint arXiv:2410.07985 , year=

work page internal anchor Pith review arXiv
[2]

Foundations and Trends

Mathematical Information Retrieval: Search and Question Answering , author=. Foundations and Trends. 2025 , publisher=

2025
[3]

Siyue Zhang, Yilun Zhao, Liyuan Geng, Arman Cohan, Anh Tuan Luu, and Chen Zhao

RaDeR: Reasoning-aware Dense Retrieval Models , author=. arXiv preprint arXiv:2505.18405 , year=

work page arXiv
[4]

Training Verifiers to Solve Math Word Problems

Training Verifiers to Solve Math Word Problems , author=. arXiv preprint arXiv:2110.14168 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Hugging Face repository , volume=

Numinamath: The largest public dataset in ai4maths with 860k pairs of competition math problems and solutions , author=. Hugging Face repository , volume=
[6]

Big-math: A large-scale, high-quality math dataset for reinforcement learning in language models.arXiv preprint arXiv:2502.17387, 2025

Big-math: A large-scale, high-quality math dataset for reinforcement learning in language models , author=. arXiv preprint arXiv:2502.17387 , year=

work page arXiv
[7]

OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems

OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems , author=. arXiv preprint arXiv:2402.14008 , year=

work page internal anchor Pith review arXiv
[8]

MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts , author=. arXiv preprint arXiv:2310.02255 , year=

work page internal anchor Pith review arXiv
[9]

2022 , howpublished=

ChatGPT , author=. 2022 , howpublished=

2022
[10]

2023 , howpublished=

GPT-4 Technical Report , author=. 2023 , howpublished=

2023
[11]

2023 , howpublished=

Claude 2 , author=. 2023 , howpublished=

2023
[12]

NeurIPS , year=

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. NeurIPS , year=
[13]

NeurIPS , year=

Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks , author=. NeurIPS , year=
[14]

2023 , howpublished=

Google Bard , author=. 2023 , howpublished=

2023
[15]

2020 , howpublished=

EasyOCR , author=. 2020 , howpublished=

2020
[16]

2023 , howpublished=

IDEFICS: An Open Multimodal Model , author=. 2023 , howpublished=

2023
[17]

2023 , howpublished=

mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality , author=. 2023 , howpublished=

2023
[18]

2023 , howpublished=

MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models , author=. 2023 , howpublished=

2023
[19]

2023 , howpublished=

LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model , author=. 2023 , howpublished=

2023
[20]

2023 , howpublished=

InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning , author=. 2023 , howpublished=

2023
[21]

2023 , howpublished=

Visual Instruction Tuning , author=. 2023 , howpublished=

2023
[22]

2023 , howpublished=

LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding , author=. 2023 , howpublished=

2023
[23]

, author=

Dense Passage Retrieval for Open-Domain Question Answering. , author=. EMNLP (1) , pages=
[24]

Unsupervised Dense Information Retrieval with Contrastive Learning

Unsupervised dense information retrieval with contrastive learning , author=. arXiv preprint arXiv:2112.09118 , year=

work page internal anchor Pith review arXiv
[25]

SIGIR , year=

ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT , author=. SIGIR , year=
[26]

arXiv preprint arXiv:2109.10086(2021)

SPLADE v2: Sparse lexical and expansion model for information retrieval , author=. arXiv preprint arXiv:2109.10086 , year=

work page arXiv
[27]

NeurIPS , year=

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. NeurIPS , year=
[28]

arXiv preprint arXiv:2007.01852 , year=

Language-agnostic BERT Sentence Embedding , author=. arXiv:2007.01852 , year=

work page arXiv 2007
[29]

NTCIR-11 , year=

NTCIR-11 Math-2 Task Overview , author=. NTCIR-11 , year=
[30]

NTCIR-12 , year=

NTCIR-12 MathIR Task Overview , author=. NTCIR-12 , year=
[31]

NTCIR-12 , year=

Tangent-3 at the NTCIR-12 MathIR Task , author=. NTCIR-12 , year=
[32]

Indexing and Searching Mathematics in Digital Libraries (MIaS) , author=
[33]

arXiv preprint arXiv:2308.13418 , year=

Nougat: Neural Optical Understanding for Academic Documents , author=. arXiv:2308.13418 , year=

work page arXiv
[34]

Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

Measuring Mathematical Problem Solving With the MATH Dataset , author=. Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=
[35]

FLoC/IJCAR , year=

MiniF2F: a cross-system benchmark for formal Olympiad-level mathematics , author=. FLoC/IJCAR , year=
[36]

GPT-4 Technical Report

Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[37]

2025 , date =

Thang Luong and Edward Lockhart , title =. 2025 , date =

2025
[38]

Measuring Multimodal Mathematical Reasoning with

Wang, Ke and Pan, Junting and Shi, Weikang and Lu, Zimu and Ren, Houxing and Zhou, Aojun and Zhan, Mingjie and Li, Hongsheng , booktitle =. Measuring Multimodal Mathematical Reasoning with
[39]

Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR , year =

SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models , author =. Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR , year =
[40]

Li, Haonan and Zhang, Yixuan and Koto, Fajri and Yang, Yifei and Zhao, Hai and Gong, Yeyun and Duan, Nan and Baldwin, Timothy , booktitle =
[41]

International Conference on Learning Representations (ICLR) , year =

Measuring Massive Multitask Language Understanding , author =. International Conference on Learning Representations (ICLR) , year =
[42]

Advances in Neural Information Processing Systems , volume=

C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models , author=. Advances in Neural Information Processing Systems , volume=
[43]

Yue, Xiang and Ni, Yuansheng and Zhang, Kai and Zheng, Tianyu and Liu, Ruoqi and Zhang, Ge and Stevens, Samuel and Jiang, Dongfu and Ren, Weiming and Sun, Yuxuan and others , booktitle =
[44]

Zhong, Wanjun and Cui, Ruixiang and Liang, Sai and others , booktitle =
[45]

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

Have llms advanced enough? a challenging problem solving benchmark for large language models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

2023
[46]

2025 , eprint=

dots.ocr: Multilingual Document Layout Parsing in a Single Vision-Language Model , author=. 2025 , eprint=

2025
[47]

2026 , eprint=

Multimodal OCR: Parse Anything from Documents , author=. 2026 , eprint=

2026
[48]

A Survey on LLM-as-a-Judge

A survey on llm-as-a-judge , author=. arXiv preprint arXiv:2411.15594 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[49]

OlympicArena: Benchmarking Multi-discipline Cognitive Reasoning for Superintelligent

Huang, Zhen and Wang, Zengzhi and Xia, Shijie and Li, Xuefeng and Zou, Haoyang and Xu, Ruijie and Fan, Run-Ze and Ye, Lyumanshan and Chern, Ethan and Ye, Yixin and others , journal =. OlympicArena: Benchmarking Multi-discipline Cognitive Reasoning for Superintelligent
[50]

Gao, Bofei and Song, Feifan and Yang, Zhe and Cai, Zefan and Miao, Yibo and Dong, Qingxiu and Li, Lei and Ma, Chenghao and Chen, Liang and Xu, Runxin and others , journal =. Omni-
[51]

arXiv preprint arXiv:2509.18383 , year=

Godel Test: Can Large Language Models Solve Easy Conjectures? , author=. arXiv preprint arXiv:2509.18383 , year=

work page arXiv
[52]

2013 , howpublished =

Gillespie, Maria , title =. 2013 , howpublished =

2013
[53]

Proceedings of the European Conference on Information Retrieval (ECIR) , year =

Advancing Math Formula Search Using Diverse Structural Features , author =. Proceedings of the European Conference on Information Retrieval (ECIR) , year =
[54]

Findings of the Association for Computational Linguistics (ACL) , year =

Reasoning in Large Language Models Through Symbolic Math Word Problems , author =. Findings of the Association for Computational Linguistics (ACL) , year =
[55]

arXiv preprint arXiv:2508.17580 , year=

UQ: Assessing Language Models on Unsolved Questions , author=. arXiv preprint arXiv:2508.17580 , year=

work page arXiv
[56]

MathArena: Evaluating LLMs on Uncontaminated Math Competitions

Matharena: Evaluating llms on uncontaminated math competitions , author=. arXiv preprint arXiv:2505.23281 , year=

work page internal anchor Pith review arXiv
[57]

Leveraging online olympiad-level math problems for llms training and contamination- resistant evaluation.CoRR, abs/2501.14275, 2025

Leveraging online olympiad-level math problems for llms training and contamination-resistant evaluation , author=. arXiv preprint arXiv:2501.14275 , year=

work page arXiv
[58]

International Conference of the Cross-Language Evaluation Forum for European Languages , pages=

Overview of ARQMath 2020: CLEF lab on answer retrieval for questions on math , author=. International Conference of the Cross-Language Evaluation Forum for European Languages , pages=. 2020 , organization=

2020
[59]

Challenging the Boundaries of Reasoning: An Olympiad-Level Math Benchmark for Large Language Models

Challenging the Boundaries of Reasoning: An Olympiad-Level Math Benchmark for Large Language Models , author=. arXiv preprint arXiv:2503.21380 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[60]

Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

Towards Robust Mathematical Reasoning , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

2025
[61]

Solving inequality proofs with large language models.CoRR, abs/2506.07927, 2025

Solving Inequality Proofs with Large Language Models , author=. arXiv preprint arXiv:2506.07927 , year=

work page arXiv
[62]

Deepseekmath-v2: Towards self-verifiable mathematical reasoning,

DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning , author=. arXiv preprint arXiv:2511.22570 , year=

work page arXiv
[63]

European Conference on Information Retrieval , pages=

Advancing Math Formula Search Using Diverse Structural and Symbolic Representations , author=. European Conference on Information Retrieval , pages=. 2025 , organization=

2025
[64]

arXiv preprint arXiv:2105.00377 , year=

Mathbert: A pre-trained model for mathematical formula understanding , author=. arXiv preprint arXiv:2105.00377 , year=

work page arXiv