The Mathematician's Assistant: Integrating AI into Research Practice

Jonas Henkel

REVIEW 3 major objections 5 minor 22 references

The Mathematician's Assistant: Integrating AI into Research Practice

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper argues that AI's current and foreseeable role for the working mathematician is augmentation, not automation: a copilot whose outputs require human verification and ownership.

desk verdict A fair, well-sourced practical survey of AI for mathematicians as of mid-2025, but its central 'augmentation not automation' claim is an extrapolation from elite competition benchmarks to ordinary research tasks, and one of its tool recommendations sits awkwardly with its own evidence. read the letter →

arxiv 2508.20236 v1 pith:NTKDHATD submitted 2025-08-27 math.HO cs.AIcs.HCcs.LG

Jonas Henkel This is my paper

classification math.HOcs.AIcs.HCcs.LG

keywords artificialintelligencelargelanguagemodelsmathematicalreasoningproofverificationbenchmarkingAI-assistedresearchhuman-AIcollaborationpractice

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

This paper makes the case that the practical, near-term value of AI for research mathematicians is augmentation, not automation. Drawing on mid-2025 benchmarks of publicly available models, it argues that today's strongest LLMs can solve hard contest problems and even grade proofs near human level, while at the same time producing plausible but invalid proofs, failing to critique their own work, and hiding uncertainty. The author therefore proposes treating AI as a copilot: the human researcher remains the pilot, responsible for direction, critical verification, and intellectual ownership. Around this idea the paper builds five guiding principles and seven concrete ways to use AI across the research lifecycle, from ideation to writing. If the picture is right, mathematicians should invest in skills like prompting, multi-model checking, and independent proof verification rather than expect AI to deliver finished results.

What carries the argument

The central object is the augmented-mathematician model, a division of labor in which the AI acts as copilot under a human pilot. Its operational core is the Principle of Critical Verification: no LLM output—proof, calculation, or summary—is accepted until independently checked. The argument is carried by two measured mechanisms from the Open Proof Corpus: the intuition gap (final-answer accuracy overstates proof validity, to a model-dependent degree) and self-critique blindness (models judge other models' proofs better than their own). Together these make a human-in-the-loop verification step the load-bearing component; best-of-n sampling appears as a practical amplifier of proof quality.

What would settle it

A direct test: take a fresh set of 100 research-level problems not present in any training corpus, have a public LLM generate complete proofs, and have independent experts or a proof assistant verify them. If a clear majority of the proofs are fully valid without human repair, the paper's 'augmentation, not automation' conclusion would be overturned.

Watch

Extended reading notes

Core claim

The paper's central claim is stated plainly in its conclusion: "the current and foreseeable impact of AI for the working mathematician is primarily one of augmentation." The evidence assembled for this is a mid-2025 snapshot of publicly accessible LLMs. On answer-only competition benchmarks such as MathArena, leading models outperform the top 1% of human contestants; on the 2025 IMO, the best accessible model scored about 31.55%, far below medal range. The Open Proof Corpus, a human-evaluated set of over 5,000 LLM-generated proofs, shows that correctness of final answers overstates proof validity: the drop from answer accuracy to full proof validity is roughly 8 percentage points for one lea

Load-bearing premise

Everything rests on treating competition-based benchmarks as faithful proxies for research-level mathematical work; if those benchmarks are contaminated or unrepresentative, the recommended balance between human and AI would be miscalibrated.

Editorial extensions

If this is right

Researchers should treat LLM proofs as drafts requiring independent verification, not as authoritative results.
Benchmarks and leaderboards that score only final answers overstate useful capability; proof-validity evaluation is needed for honest comparisons.
Because self-critique blindness is strongest on a model's own proofs, a practical workflow is to generate with one model and verify with another.
Best-of-n sampling is a cheap way to raise proof success rates; the paper reports one model improving from 26% to 47% by choosing the best of eight ranked proofs.
Graduate training should add AI collaboration skills—prompting, critical evaluation, and ethical use—alongside conventional mathematical training.

Reading between the lines

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

The answer-vs-proof gap suggests a division of labor the paper only partly spells out: use high-answer-accuracy models for conjecture generation, then switch to verification-oriented tools for proof checking.
If proof-validity benchmarks become standard and models keep improving, the 'augmentation' verdict may shift task by task—automation may arrive first in optimization-style problems and last in open-ended proof construction.
A testable extension is a two-model pipeline (generator plus independent judge) on a fresh problem set; the Open Proof Corpus data predicts this will beat same-model self-checking.
The self-critique blindness finding implies that any claim of 'AI verified its own proof' should be treated skeptically; independent verification is the only meaningful check.

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.

Desk Editor's Note

read the letter

Likely worth a read if you are writing about AI and mathematical practice, less because it breaks new ground than because it is a fair, well-sourced snapshot of where things stood at the start of August 2025. The 'augmented mathematician' framing is not new, and the seven usage categories are standard, but the paper did the work of pulling together MathArena, the Open Proof Corpus, and FrontierMath data and it reports them mostly faithfully. It also deserves credit for flagging the FrontierMath funding controversy and for not over-claiming OpenAI's internal results.

The soft spot is exactly where the stress-test put it. Section 6 concludes that AI's impact on the working mathematician is 'primarily augmentation,' and that recommendation is calibrated by low scores on deliberately elite problems (IMO ~30%, FrontierMath Tier 4 ~4%). But the Seven Ways are mostly routine tasks: literature search, code generation, notation consistency, writing. The paper gives no evidence about how much of those tasks can be automated, so the augmentation conclusion is extrapolated from the hardest proof-completion problems. The Ken Ono and Terence Tao quotes actually underscore that those problems are at the edge of what professional mathematicians can do, so a low model score there says very little about the rest of the workflow. That does not sink the paper, but it means the central claim is an opinion, not a result.

A second, smaller issue: the paper relies on vendor and press claims (IMO gold for Deep Think, AlphaEvolve's 20% 'new solutions') without independent verification. It explicitly says it is being conservative with FrontierMath because of the OpenAI funding, but it does not apply the same standard to DeepMind's self-reported results.

Third, the 'third-party arbiter' suggestion in Section 5.6.2 sits oddly with the self-critique data in Section 3.2.3. If models are notably bad at judging their own proofs, then a 'trusted model' used to settle a disagreement between two humans is not obviously trustworthy; the OPC numbers show average judgment accuracy near human level, but the diagonal effect is a real limitation the paper itself documents. A multi-model cross-check would be more consistent with the evidence.

None of these issues make the paper a waste of time. It is a useful, honest practical guide, and it is explicit about its limits. For a journal that publishes expository work on mathematical practice, I would send it to review with a request to address the benchmark-to-research gap and to moderate the arbiter claim. For a research journal, it is not a research contribution.

Referee Report

3 major / 5 minor

Summary. The paper is a practice-oriented survey of AI tools for mathematical research as of August 2025. It summarizes benchmark evidence from MathArena, the Open Proof Corpus, and FrontierMath, describes landmark claims such as Gemini Deep Think's IMO gold medal and AlphaEvolve's algorithmic discoveries, and proposes a framework built on five guiding principles (e.g., copilot-not-pilot, critical verification) and seven areas of AI use across the research lifecycle. Its central conclusion, stated in §6, is that AI currently and foreseeably serves primarily as an augmentation of the working mathematician rather than as an automation of mathematical research, and that researchers therefore need skills in prompting, verification, and methodological rigor. The paper is explicitly scoped to publicly accessible models as of August 2, 2025, and does not claim to provide original empirical measurements.

Significance. If taken as a synthesis and practical guide, the paper is useful and timely. It collects relevant benchmark results, correctly emphasizes the distinction between final-answer accuracy and proof validity, highlights the self-critique blindness documented in the Open Proof Corpus, and gives concrete, cautious advice on verification and model selection. It also handles the FrontierMath funding controversy more transparently than many popular accounts. However, the paper is not an original empirical study: nearly all of its evidence comes from vendor reports, preprints, and press coverage. The main risk is that the broad conclusion in §6—'augmentation rather than automation' for the whole research workflow—is inferred from benchmarks that measure only a narrow slice of mathematical work, namely elite competition problems and proof-completion tasks. The practical framework may be reasonable, but the evidentiary support for the general claim is incomplete.

major comments (3)

[§6, §3.4] The central conclusion that AI's impact on the working mathematician is 'primarily one of augmentation' is not supported for the full research workflow by the cited evidence. The benchmarks discussed—IMO, USAMO, FrontierMath—target deliberately difficult proof-completion problems. Section 3.4 itself quotes Ken Ono and Terence Tao saying that Tier-4 FrontierMath problems were designed to overwhelm AI and may require a semi-expert plus AI. A 4.2% success rate on such problems says little about literature search, routine lemma verification, code generation, counterexample search, or formalization, which are the tasks emphasized in §5.1–5.7. To make the claim stand, the paper would need a task-level analysis or a distributional argument over actual research activities, or it should explicitly restrict the conclusion to proof-generation tasks. This is load-bearing because §6 states the conclu
[§2, §3.3] Several landmark claims are treated as established facts without the same caution applied to FrontierMath in §3.4. The IMO gold medal of Gemini Deep Think, AlphaEvolve's 20% success rate, and the 4×4 matrix-multiplication improvement all come from vendor announcements, blog posts, or preprints, not from independent evaluation. Section 2 calls the IMO result 'a landmark achievement' and builds part of the paper's calibration on it, but the only cited sources are DeepMind, Reuters, and a technical report. Since §6 uses these results to argue that AI capability is 'rapidly maturing,' the paper should explicitly label all vendor-reported claims as such, state whether independent verification exists, and qualify the conclusions that depend on them.
[§3.2.2, §5.1.1] The interpretation of the OPC 'intuition gap' goes beyond what the data show. A gap between final-answer accuracy and proof validity does not establish that models possess mathematical intuition; it may reflect answer-only heuristics, partial pattern matching, or accidental correctness. Section 5.1.1 then uses this interpretation to recommend using a high-answer-accuracy model to generate conjectures. That is a reasonable heuristic, but it should be presented as an inference from the benchmark, not as a demonstrated model property. Similarly, the multi-model 'peer review' recommendation in §5.5.1 rests on anecdotal experience; if it is a core recommendation, it should be labeled as currently unvalidated.

minor comments (5)

[Abstract, §3.2.2] The phrase 'a model depending discrepancy' appears twice in the abstract; this should be 'a model-dependent discrepancy.'
[§5.5.2] There is a grammatical error: 'may transforms to its strongest applications' should read 'may be one of its strongest applications' or similar.
[§4.2.1, Table 1] The price is given as '€275/month' in the text and '€274.99/month' in the table; this is a minor consistency issue.
[§4.2.4, Table 3] The naming of the highest-tier model is inconsistent: 'Grok-4-Heavy' in the table and 'Grok-4-Heavy' in the text; also the table calls it 'SuperGrok Heavy.' Please harmonize.
[§5.8] The appeal to Dean and Naibo for 'theoretical arguments' that mathematical problems are inherently resistant to algorithmic resolution cites only a Substack summary. The original publication or a more precise statement of the argument should be provided if this philosophical point is retained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's empirical claims are anchored in external benchmarks and its normative framework is not presented as a derivation from those inputs.

full rationale

The paper makes no mathematical derivation claims. Its empirical sections (Sections 2–3) compile externally reported results from MathArena, the Open Proof Corpus, the 2025 IMO, and FrontierMath, all cited to independent preprints and datasets. No parameter is fitted to a subset of data and then relabeled as a prediction; no quantity is defined in terms of the conclusion it is supposed to support; and no uniqueness theorem or load-bearing ansatz is imported from the author's own prior work. The central conclusion in Section 6—that AI's current and foreseeable role is primarily augmentation—is a synthetic judgment informed by those external benchmarks, not a result forced by construction. The Five Guiding Principles and Seven Ways are explicitly framed as a proposed practical framework, not as consequences derived from the benchmark numbers. The paper even exercises caution by setting aside OpenAI's FrontierMath results because of contamination and funding concerns, which runs counter to a circular reliance on vendor-reported claims. The only self-referential element is the Acknowledgements note that AI tools were used in writing the text; this is a transparency disclosure, not a load-bearing premise. Concerns about benchmark-to-research extrapolation, which are legitimate, concern external validity and calibration, not circularity. Therefore there are no circular steps and the score is 0.

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

This is a review/position paper, so the ledger is small: no free parameters, no invented entities. The central recommendations depend on three domain assumptions about benchmarks representing research mathematics, and on the assumption that human researchers can reliably perform critical verification of LLM output.

assumptions (4)

domain assumption IMO and similar contests are valid measures of creative mathematical reasoning.
Section 2: the paper calls the IMO 'an exceptional benchmark not for accumulated knowledge, but for genuine, creative problem-solving ability.' The augmentation conclusion leans on this.
domain assumption MathArena and Open Proof Corpus results are uncontaminated and accurately reported.
Sections 3.1-3.2: the paper's capability claims rely on these benchmarks, including the 43% overall proof correctness and the 30% final-answer/proof gap for o3. The paper does not re-verify the source data.
domain assumption Competition performance transfers to professional research mathematics.
Sections 3.4 and 6: the paper extrapolates from competition scores to research practice, arguing that current AI is an assistant rather than an autonomous researcher. If FrontierMath Tier 4's 4.2% is the true research-level ceiling, that supports augmentation; but this is one narrow benchmark.
domain assumption Humans can reliably verify AI-generated proofs.
Section 4.1 Principle 2 prescribes critical verification but provides no evidence about human accuracy in detecting LLM proof errors, which is central to the safety of the copilot model.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Mathematician's Assistant: Integrating AI into Research Practice." pith.science (2026). https://pith.science/paper/NTKDHATD

@misc{pith2026250820236,
  author       = {Pith},
  title        = {Pith review of: The Mathematician's Assistant: Integrating AI into Research Practice},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NTKDHATD}},
  note         = {Machine review of arXiv:2508.20236}
}

read the original abstract

The rapid development of artificial intelligence (AI), marked by breakthroughs like 'AlphaEvolve' and 'Gemini Deep Think', is beginning to offer powerful new tools that have the potential to significantly alter the research practice in many areas of mathematics. This paper explores the current landscape of publicly accessible large language models (LLMs) in a mathematical research context, based on developments up to August 2, 2025. Our analysis of recent benchmarks, such as MathArena and the Open Proof Corpus (Balunovi\'c et al., 2025; Dekoninck et al., 2025), reveals a complex duality: while state-of-the-art models demonstrate strong abilities in solving problems and evaluating proofs, they also exhibit systematic flaws, including a lack of self-critique and a model depending discrepancy between final-answer accuracy and full-proof validity. Based on these findings, we propose a durable framework for integrating AI into the research workflow, centered on the principle of the augmented mathematician. In this model, the AI functions as a copilot under the critical guidance of the human researcher, an approach distilled into five guiding principles for effective and responsible use. We then systematically explore seven fundamental ways AI can be applied across the research lifecycle, from creativity and ideation to the final writing process, demonstrating how these principles translate into concrete practice. We conclude that the primary role of AI is currently augmentation rather than automation. This requires a new skill set focused on strategic prompting, critical verification, and methodological rigor in order to effectively use these powerful tools.

Figures

Figures reproduced from arXiv: 2508.20236 by the authors.

**Figure 2.** MathArena: Main results on numerical answer evaluation on competitions between January-May 2025. Green cells denote models released after the respective competition date [Ba+25, Tab. 2]. 3.1.2. Proof-Based Performance. The assessment of LLM capabilities becomes more nuanced when shifting to proof-based evaluations. This format requires a definitive answer and a complete, step-by-step logical reasoning process. Thi… view at source ↗

**Figure 3.** Average proof correctness of various models on the OPC. Data is split into two partitions. First, resp. second was answered by all models except for DeepSeek-R1, resp. Grok 3 mini [De+25, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Comparison of final-answer accuracy and proof correctness on the MathArena subset [De+25, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Benchmarking LLMs as proof graders. Costs are given in USD [De+25, Tab. 2] [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Judgement accuracy breakdown split by solver, highlighting the lowest score for each judge [De+25, Tab. 3] [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Performance of leading LLMs on the 50 problems of FrontierMath’s Tier 4, the most difficult subset of the benchmark [EpAi25] [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

[1]

et al.\ (2025)

Balog, M. et al.\ (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131 https://arxiv.org/abs/2506.13131

arXiv 2025
[2]

et al.\ (2025)

Balunovi\' c , M. et al.\ (2025). MathArena: Evaluating LLMs on Uncontaminated Math Competitions. arXiv preprint arXiv:2505.23281 https://arxiv.org/abs/2505.23281

arXiv 2025
[3]

et al.\ (2025)

Balunovi\' c , M. et al.\ (2025). MathArena: Evaluating LLMs on Uncontaminated Math Competitions. SRI Lab, ETH Zurich. Available at: matharena.ai https://matharena.ai/

work page 2025
[4]

et al.\ (2025)

Balunovi\' c , M. et al.\ (2025). Not Even Bronze: Evaluating LLMs on 2025 International Math Olympiad. MathArena Blog (SRI Lab, ETH Zurich), 25 Jul 2025. Available at: matharena.ai https://matharena.ai/imo/

work page 2025
[5]

et al.\ (2024)

Barkley, G. et al.\ (2024). FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI. arXiv preprint arXiv:2411.04872 https://arxiv.org/abs/2411.04872

arXiv 2024
[6]

Google clinches milestone gold at global math competition, while OpenAI also claims win

Cai, K., Singh, J.\ (2025). Google clinches milestone gold at global math competition, while OpenAI also claims win. Reuters, 22 Jul 2025. Available at: reuters.com https://www.reuters.com/world/asia-pacific/google-clinches-milestone-gold-global-math-competition-while-openai-also-claims-2025-07-22/

work page 2025
[7]

Dean, W., Naibo, A. (2025). Artificial Intelligence and Inherent Mathematical Difficulty (summary). Silicon Reckoner (Substack), 22 Jun 2025. Available at: siliconreckoner.substack.com https://siliconreckoner.substack.com/p/artificial-intelligence-and-inherent

work page 2025
[8]

et al.\ (2025)

Dekoninck, J. et al.\ (2025). The Open Proof Corpus: A Large-Scale Study of LLM-Generated Mathematical Proofs. arXiv preprint arXiv:2506.21621 https://arxiv.org/abs/2506.21621

arXiv 2025

Show all 22 references

[9]

Eadicicco, L., Morrow, A. (2025). Grok's antisemitic outbursts reflect a problem with AI chatbots. CNN, 10 Jul 2025. Available at: cnn.com https://edition.cnn.com/2025/07/10/tech/grok-antisemitic-outbursts-reflect-a-problem-with-ai-chatbots

2025
[10]

FrontierMath — benchmarking AI against advanced mathematical research Available at: epoch.ai https://epoch.ai/frontiermath

Epoch AI (2025). FrontierMath — benchmarking AI against advanced mathematical research Available at: epoch.ai https://epoch.ai/frontiermath

2025
[11]

About FrontierMath

Epoch AI (2025). About FrontierMath. Available at: epoch.ai https://epoch.ai/frontiermath/about

2025
[12]

Clarifying the Creation and Use of the FrontierMath Benchmark

Epoch AI (2025). Clarifying the Creation and Use of the FrontierMath Benchmark. Epoch AI Blog, 23 Jan 2025. Available at: epoch.ai https://epoch.ai/blog/openai-and-frontiermath

2025
[13]

Generative AI: Hype, Or Truly Transformative? Goldman Sachs Global Macro Research, Issue 120

Goldman Sachs (2023). Generative AI: Hype, Or Truly Transformative? Goldman Sachs Global Macro Research, Issue 120

2023
[14]

AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms

Google DeepMind (2025). AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms. DeepMind Blog, 14 May 2025. Available at: deepmind.google https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/

2025
[15]

Gemini Deep Think for International Mathematical Olympiad 2025

Google DeepMind (2025). Gemini Deep Think for International Mathematical Olympiad 2025. Technical report, July 2025. Available at: storage.googleapis.com https://storage.googleapis.com/deepmind-media/gemini/IMO_2025.pdf

2025
[16]

The FrontierMath scandal

Harris, M.\ (2025). The FrontierMath scandal. Silicon Reckoner (Blog), 26 Jan 2025. Available at: siliconreckoner.substack.com https://siliconreckoner.substack.com/p/the-frontier-math-scandal

2025
[17]

Advanced version of Gemini with Deep Think officially achieves gold‑medal standard at the International Mathematical Olympiad

Lockhart, E., Luong, T.\ (2025). Advanced version of Gemini with Deep Think officially achieves gold‑medal standard at the International Mathematical Olympiad. DeepMind Blog, 21 Jul 2025. Available at: deepmind.google https://deepmind.google/discover/blog/advanced-version-of-g...

2025
[18]

International Mathematical Olympiad (IMO) 2025 Leaderboard

MathArena (2025). International Mathematical Olympiad (IMO) 2025 Leaderboard. Available at: matharena.ai/imo https://matharena.ai/imo/

2025
[19]

et al.\ (2025)

Maslej, N. et al.\ (2025). Artificial Intelligence Index Report 2025. Stanford, CA: Stanford University

2025
[20]

et al.\ (2025)

Petrov, I. et al.\ (2025). PROOF OR BLUFF? EVALUATING LLMS ON 2025 USA MATH OLYMPIAD. arXiv preprint arXiv:2503.21934v4 https://arxiv.org/abs/2503.21934

2025 arXiv
[21]

Embracing Change and Resetting Expectations

Tao, T.\ (2023). Embracing Change and Resetting Expectations. In: E. Horvitz (Ed.), AI Anthology. Available at: unlocked.microsoft.com https://unlocked.microsoft.com/ai-anthology/terence-tao/

2023
[22]

et al.\ (2025)

Zanibbi, R. et al.\ (2025). Mathematical Information Retrieval: Search and Question Answering. arXiv preprint arXiv:2408.11646 https://arxiv.org/abs/2408.11646

2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

et al.\ (2025)

Balog, M. et al.\ (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131 https://arxiv.org/abs/2506.13131

arXiv 2025

[2] [2]

et al.\ (2025)

Balunovi\' c , M. et al.\ (2025). MathArena: Evaluating LLMs on Uncontaminated Math Competitions. arXiv preprint arXiv:2505.23281 https://arxiv.org/abs/2505.23281

arXiv 2025

[3] [3]

et al.\ (2025)

Balunovi\' c , M. et al.\ (2025). MathArena: Evaluating LLMs on Uncontaminated Math Competitions. SRI Lab, ETH Zurich. Available at: matharena.ai https://matharena.ai/

work page 2025

[4] [4]

et al.\ (2025)

Balunovi\' c , M. et al.\ (2025). Not Even Bronze: Evaluating LLMs on 2025 International Math Olympiad. MathArena Blog (SRI Lab, ETH Zurich), 25 Jul 2025. Available at: matharena.ai https://matharena.ai/imo/

work page 2025

[5] [5]

et al.\ (2024)

Barkley, G. et al.\ (2024). FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI. arXiv preprint arXiv:2411.04872 https://arxiv.org/abs/2411.04872

arXiv 2024

[6] [6]

Google clinches milestone gold at global math competition, while OpenAI also claims win

Cai, K., Singh, J.\ (2025). Google clinches milestone gold at global math competition, while OpenAI also claims win. Reuters, 22 Jul 2025. Available at: reuters.com https://www.reuters.com/world/asia-pacific/google-clinches-milestone-gold-global-math-competition-while-openai-also-claims-2025-07-22/

work page 2025

[7] [7]

Dean, W., Naibo, A. (2025). Artificial Intelligence and Inherent Mathematical Difficulty (summary). Silicon Reckoner (Substack), 22 Jun 2025. Available at: siliconreckoner.substack.com https://siliconreckoner.substack.com/p/artificial-intelligence-and-inherent

work page 2025

[8] [8]

et al.\ (2025)

Dekoninck, J. et al.\ (2025). The Open Proof Corpus: A Large-Scale Study of LLM-Generated Mathematical Proofs. arXiv preprint arXiv:2506.21621 https://arxiv.org/abs/2506.21621

arXiv 2025