Pith. sign in

REVIEW 3 major objections 8 minor 60 references

Giving language models multi-turn SageMath access raises solve rates on research-level computational math by about 10 points on average.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 20:37 UTC pith:SHGU4DBN

load-bearing objection Clean controlled ablation: SageMath access lifts every model on a refined RealMath subset, with the real value in the trace-level behavioral analysis rather than the headline +9.7 pp. the 3 major comments →

arxiv 2607.06820 v1 pith:SHGU4DBN submitted 2026-07-07 cs.AI

Evaluating SageMath-Augmented LLM Agents for Computational and Experimental Mathematics

Pavel Snopov , German Magai This is my paper
classification cs.AI MSC 68T0568W3000A35
keywords LLM agentscomputer algebra systemsSageMathReActresearch-level mathematicsRealMath benchmarktool-augmented reasoningconjecture discovery
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether verifiable symbolic feedback from a computer algebra system helps frontier language models on research-level mathematics, not just competition problems or formal proofs. The authors put fifteen models through a matched ablation: pure reasoning versus a ReAct-style loop that can call SageMath and look up its documentation. On a curated set of 133 RealMath problems with checkable numerical or symbolic answers, every model improves; the average gain is 9.7 percentage points, with open-weight models often gaining the most. The best tool-enabled run reaches a 75.2% solve rate at the lowest token cost among those setups. The authors treat this as evidence that CAS-augmented agents can assist computational exploration and edge toward automated conjecture discovery.

Core claim

Across fifteen frontier models on 133 research-level problems, multi-turn access to SageMath (with documentation retrieval) improves solve rate for every model, by +9.7 percentage points on average (range +1.5 to +27.8). Open-weight models gain more than closed ones on average and close much of the gap; GPT-5.5 with tools reaches the highest solve rate, 75.2%, while also using the fewest tokens among tool-enabled configurations.

What carries the argument

A controlled ReAct-style agentic loop with multi-turn SageMath execution and Context7 documentation retrieval, ablated against a tool-free baseline on the same models and problems, with answers scored by a hybrid pipeline of SymPy symbolic equivalence plus majority-vote LLM judges.

Load-bearing premise

That the filtered 133-problem set of numerically or symbolically checkable, SageMath-feasible tasks, and the hybrid automatic-plus-LLM validation of answers, are fair enough that the measured gains generalize to real computational research mathematics.

What would settle it

Rerun the same tool-access ablation on a larger hold-out set of research problems whose answers are independently verified by human mathematicians (or a fully deterministic symbolic checker), and check whether average solve-rate gains remain near +10 points and whether open-weight models still close most of the gap.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • CAS feedback is a practical, model-agnostic lever for research-level computational math, not only for competition-style or formal-proof tasks.
  • Open-weight models can approach closed-model tool-free performance once they can iterate against SageMath.
  • Token spend does not track accuracy under tool use: efficient short verification loops beat long brute-force tool traces.
  • Agent traces that recover closed forms by computing intermediates, spotting patterns, and testing conjectures support CAS agents as tools for experimental mathematics.
  • Refined RealMath-style extraction plus hybrid symbolic/LLM validation is a reusable protocol for executable research-math benchmarks.

Where Pith is reading between the lines

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

  • Training or scaffolding models specifically for short verify-and-recover tool loops may matter more than raw tool-call volume.
  • Problems that stay unsolved even with CAS access likely need deeper theory or specialized knowledge beyond current agent loops.
  • A multi-agent split—hypothesis generation, CAS experimentation, and independent checking—could amplify the conjecture-discovery pattern shown in the case study.
  • Domains with mature CAS libraries (combinatorics, algebra) should show larger tool gains than areas still poorly formalized in software.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. The paper studies whether multi-turn access to SageMath (plus Context7 documentation) improves frontier LLMs on research-level computational mathematics. It evaluates 15 models on a curated 133-problem subset of RealMath under a matched tool-free vs ReAct-style tool-augmented ablation, with a hybrid SymPy + three-model LLM-judge validation pipeline and extensive tool-trace analysis. The headline empirical result is a uniform solve-rate gain under tool access (+9.7 pp on average; range +1.5 to +27.8 pp), with GPT-5.5 reaching 75.2% at the lowest tool-enabled token cost, open-weight models gaining more than closed ones, and behavioral patterns (bimodal traces, recovery after failures, category-wise effects) characterizing effective vs ineffective CAS use. A case study illustrates iterative example computation, pattern extraction, and formula conjecture for torsion orders of twisted torus knots. The authors also refine RealMath via LLM-assisted filtering, context compression, and SymPy-normalizable answer extraction, and release code.

Significance. The work addresses a genuine gap: AI4Math has emphasized autoformalization and theorem proving far more than CAS-based experimental workflows that practicing computational mathematicians actually use. The controlled tool-access ablation, multi-model scope, hybrid validation with partial human audit, and unusually detailed error/trace analysis (exception types, recovery rates, bimodal turn distributions, function usage by arXiv category) make the empirical contribution more than a leaderboard dump. The case study is a concrete demonstration that agents can use SageMath for intermediate objects rather than as an oracle. Strengths that should be credited include the matched ablation design, Wilson intervals, public repository, and an honest Limitations section that flags selection bias and confounds. If the measured gains hold under the stated protocol, the paper is a solid step toward CAS-augmented agents for computational exploration.

major comments (3)
  1. §3.4 and the validation-stage analysis in §4.1: the pipeline routes 60.8% of 3,990 predictions to a three-judge panel and states that 30% of those cases were manually inspected and treated as ground truth, yet the manuscript never reports human–panel agreement (or human overturn rate). Because absolute solve rates and small inter-model gaps rest on this stage (the authors themselves note that the judge recovers 747 answers and that small gaps are methodologically fragile; §4.3), please report agreement statistics between the majority vote and the human annotations, stratified if possible by answer type (numerical vs expression). Without that number, the precision of the 75.2% / +9.7 pp headlines is harder to assess than the design otherwise warrants.
  2. §A.2 Dataset Construction and §4.3 Limitations vs Abstract/Conclusions: the evaluation set is obtained by aggressive LLM-assisted filtering for SymPy-normalizable answers and for SageMath-feasibility (633 → 300 → 219 → 133). That is appropriate for studying CAS-augmented agents, but it means measured gains are on problems pre-selected as computationally executable. The abstract and §6 still frame results as gains on “research-level mathematical problems” and as extending practical reasoning capabilities more broadly. Please tighten the claim scope in the abstract and conclusions so that the primary result is explicitly “on a SageMath-feasible, expression/numerical subset of RealMath,” and keep the more general language as a carefully caveated interpretation rather than the headline.
  3. §4.3 Limitations correctly states that the tool-augmented condition combines SageMath, Context7, ReAct prompting, and a 15-call budget, so gains cannot be attributed to CAS access alone; Finding 4 further shows that most models barely use Context7. The abstract, title emphasis, and Finding 1 nonetheless attribute improvements primarily to “SageMath access.” Either add a brief Context7-off or documentation-ablated control for at least one strong and one weak model, or rephrase the abstract/findings to “CAS-and-documentation tool package under ReAct” while noting that Context7 usage is negligible for most models. As written, the causal language is slightly stronger than the ablation isolates.
minor comments (8)
  1. §5 title and §6: “Conjecture Discovery” / “step towards automated conjecture discovery” overstates the case study. The agent rediscovers closed forms already present in Himeno & Teragaito (2024) via finite-instance Sage evidence plus known Floer-theoretic structure; the manuscript itself says the trace “provides finite-instance evidence rather than a proof.” Reframe as “computational rediscovery / conjecture formulation on known results” unless new open conjectures are produced.
  2. Figure 2 / Figure 3 captions and axis labels: ensure solve-rate percentages and token scales are fully legible in grayscale and that Wilson intervals are defined once in the caption as well as in §3.4.
  3. Introduction: “and and have recently autonomously resolved” — duplicate “and”; clean residual typos of this kind throughout.
  4. §3.1 / Appendix A.2: state more clearly how many problems were discarded at each filter stage for each reason (non-SymPy-normalizable vs not Sage-feasible vs sampling for category balance), ideally as a small flowchart or table, so the selection funnel is auditable without reading the full appendix narrative.
  5. Related Work §2.2: IMPROOFBENCH and ASyMOB comparisons are useful; a one-sentence explicit contrast table (problem count, public/private, tool ablation present/absent, answer type) would help readers place RealMath-133 quickly.
  6. §4.1 “descried by incorrect tool interaction” appears to be a typo for “described” or “derailed.”
  7. Prompts in Appendix F are a reproducibility strength; consider noting temperature / decoding settings and API dates for each model family in §3.3 or A.1, since frontier APIs drift.
  8. Figure 5 and the “11 problems solved by no configuration” discussion: a short qualitative characterization of those unsolved items (beyond “deep theoretical insight”) would strengthen the claim that CAS feedback is insufficient there.

Circularity Check

0 steps flagged

No circular derivation: central claims are controlled empirical ablations of solve rate under tool access, not quantities forced by definition or self-citation.

full rationale

The paper’s load-bearing chain is empirical, not definitional. It fixes a curated 133-problem subset of RealMath, holds model and problem fixed, and compares tool-free direct answering against a ReAct loop with SageMath (+Context7) under a 15-call budget; solve rate is accepted by SymPy difference-to-zero or, when inconclusive, majority vote of three LLM judges (with partial human audit). Reported gains (+9.7 pp average; range +1.5 to +27.8 pp; GPT-5.5 at 75.2%) are measured outcomes of that ablation, not parameters fitted to data and re-labeled as predictions, nor results forced by uniqueness theorems or self-citations from the present authors. Dataset filtering for SymPy-normalizable / Sage-feasible items and residual judge subjectivity are selection and measurement risks already stated in Limitations §4.3 and Appendix A.2; they do not make the measured deltas true by construction. The case study (torsion orders of twisted torus knots) illustrates an agent recovering known closed forms via intermediate Sage computations; it is presented as workflow evidence, not as a first-principles derivation of a new theorem. No self-definitional loop, fitted-input-as-prediction, load-bearing self-citation, or renaming of a known result as a novel derivation is present. Honest finding: no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

Empirical evaluation paper; load-bearing premises are standard experimental-design choices and the reliability of the external tools rather than novel physical or mathematical postulates.

free parameters (2)
  • max_tool_calls
    Hard cap of 15 Sage/Context7 calls per problem; directly shapes the observed bimodal trace-length distribution and the forced-finalization rate of weaker models.
  • sandbox_wall_clock_timeout
    45-second limit per Sage execution; determines which enumerative strategies are classified as timeouts versus productive computation.
axioms (3)
  • domain assumption SageMath (and its underlying GAP/Singular/PARI engines) correctly implements the algebraic and combinatorial operations invoked by the agents.
    All verifiable feedback rests on the correctness of the CAS; the paper treats it as an oracle.
  • domain assumption Two SymPy expressions that simplify to a zero difference, or that a majority of three frontier LLM judges deem equivalent, are mathematically the same answer for scoring purposes.
    The hybrid validation pipeline is the sole source of the solve-rate metric.
  • ad hoc to paper The ReAct loop with the supplied system prompts is a reasonable emulation of a computational mathematician’s iterative experiment–conjecture cycle.
    Used to justify the claim that the setup advances automated conjecture discovery.

pith-pipeline@v1.1.0-grok45 · 34866 in / 2382 out tokens · 38839 ms · 2026-07-10T20:37:58.913287+00:00 · methodology

0 comments
read the original abstract

Recent advances in AI for Mathematics have focused largely on autoformalization and theorem proving, leaving the role of Computer Algebra Systems (CAS) in agentic LLM workflows underexplored. We propose a ReAct-style agentic setup that combines LLM reasoning with verifiable feedback from SageMath, together with Context7 for the up-to-date documentation. We evaluate this agentic setup across frontier models for solving research-level mathematical problems from the RealMath benchmark in a setting that emulates a computational-mathematics research loop. We also propose a refinement to the RealMath benchmark by introducing a multi-step post-processing procedure and a multi-stage validation pipeline, both of which improve the quality and reliability of the extracted problem set. Our experiments reveal substantial performance gains from SageMath access across all evaluated models on +9.7~pp on average, the gains range from 1.5~pp to 27.8~pp and narrow the gap between open-weight and closed models. Qwen~3.7-Max benefits from SageMath the most, while GPT-5.5 achieves the highest solve rate of $75.2\%$ and the lowest token usage among tool-enabled configurations. Our findings suggest that CAS-augmented agents represent a promising direction for assisting mathematicians in computational exploration, and we believe that this work is a step towards automated conjecture discovery. The project repository is available online.

Figures

Figures reproduced from arXiv: 2607.06820 by German Magai, Pavel Snopov.

Figure 1
Figure 1. Figure 1: Overview of the proposed evaluation framework for executable mathematical reasoning. A ReAct-style LLM agent interacts with SageMath and Context7 through iterative tool use, receives verifiable feedback from symbolic computation, and produces a final answer validated through symbolic equivalence checking and an auxiliary LLM-as-a-Judge stage. 3.2. Agentic Setup The agentic pipeline is organized to emulate … view at source ↗
Figure 2
Figure 2. Figure 2: Solve rate across models and evaluation setups. Solve rates for each model without tools and with the tool [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Solve rates against mean token usage per problem on a logarithmic axis, showing the accuracy-cost trade-off. The largest multipliers (54–99×; Qwen, DeepSeek 3.2, and DeepSeek V4-flash) belong to the models with the lowest tool-free token usage (2.3–2.5k tokens per problem), so the multiplier largely reflects the terseness of the tool-free model rather than the cost of the agentic loop itself. GPT-5.5 attai… view at source ↗
Figure 4
Figure 4. Figure 4: Mean tokens per problem on the common all-solved set in both tool-free and tool-enabled setups. Model efficiency on commonly solved problems. In practice, one often faces the choice of which model to de￾ploy when a problem is relatively easy and the goal is to solve it at minimal cost. Here we therefore focus to the subset of problems solved by model within a given setup, accuracy is tied on this set by co… view at source ↗
Figure 5
Figure 5. Figure 5: Problem-level solve coverage across configurations. Each bar shows number of problems solved by N configurations. sole exception with a slightly higher expression score (68% vs. 67%). Tool access narrows this disparity where it is widest: Qwen gains +36 pp on expressions (27% → 63%) and DeepSeek 3.2 +18 pp, whereas models already strong on expressions (GPT-5.5, Fugu-Ultra, Gemini 3.1 Pro) gain only 1–5 pp.… view at source ↗
Figure 6
Figure 6. Figure 6: Solve rate by arXiv category with and without tools. Cells show solve rate and solved/total counts. Solve rate by mathematical category. We further exam￾ine the effect of the agentic setup on performance across arXiv categories ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Distribution of agentic trace lengths across models. Each bar shows the number of problems for which the agent terminated after a given number of tool calls. Tool-use depth [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Mean SageMath sandbox runtime per problem in the tool-augmented setting, with the sandbox share of total solve time. are often dominated by repeated successful calls, whereas Kimi’s contain dense mixtures of code errors and timeouts. Grok 4.3 exhibits a more polarized pattern, alternating be￾tween very short successful traces and long clusters of code errors. Overall, the visualization suggests that strong… view at source ↗
Figure 9
Figure 9. Figure 9: The relationship between the mean number of tool calls and the solve rate for various arXiv categories [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Pairwise per-problem agreement matrices across all evaluated models. Left: tool-free setting. Right: agentic setting with tools. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Solve rate by answer type for each model and setup. Bars compare expression and numerical problems under no-tool and tool-augmented runs [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Tool calls by model. Bars show the total number of Sage execution calls and Context7 lookups across benchmark problems [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Failed Sage calls broken down by Python exception type (left) and by model (right). 19 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Recovery after an error per model. C. Function usage by arXiv category [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Tool-call traces for the selected 50 problems across models. Each column is one problem and each vertical stack shows the sequence of tool calls made by a model in a single run. Colors indicate successful Sage calls, code errors, timeouts, runtime crashes, documentation queries, missing tool calls. The small bar below each stack marks the final answer as correct or incorrect. 21 [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 16
Figure 16. Figure 16: Function-level tool usage by arXiv primary category. Rows show the top functions extracted from tool traces, after filtering constructors and generic programming utilities. Columns show arXiv primary categories. Each cell reports the number of Sage execution traces in which the function appears at least once. These labels are normalized and indicate the mathematical subdomain and function or method name. … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages · 14 internal anchors

  1. [1]

    and Li, Wenda and Rabe, Markus N

    Wu, Yuhuai and Jiang, Albert Q. and Li, Wenda and Rabe, Markus N. and Staats, Charles and Jamnik, Mateja and Szegedy, Christian , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  2. [2]

    Goedel-Prover-V2: Scaling Formal Theorem Proving with Scaffolded Data Synthesis and Self-Correction

    Lin, Yong and Tang, Shange and Lyu, Bohan and Yang, Ziran and Chung, Jui-Hui and Zhao, Haoyu and Jiang, Lai and Geng, Yihan and Ge, Jiawei and Sun, Jingruo and Wu, Jiayun and Gesi, Jiri and Lu, Ximing and Acuna, David and Yang, Kaiyu and Lin, Hongzhou and Choi, Yejin and Chen, Danqi and Arora, Sanjeev and Jin, Chi , title =. arXiv preprint arXiv:2508.0361...

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Aristotle: IMO-level Automated Theorem Proving

    Achim, Tudor and Best, Alex and Bietti, Alberto and Der, Kevin and F. arXiv preprint arXiv:2510.01346 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2601.07421 , year =

    Sothanaphan, Nat , title =. arXiv preprint arXiv:2601.07421 , year =

    work page arXiv

  5. [5]

    and Stillman, Michael E

    Grayson, Daniel R. and Stillman, Michael E. , title =

    work page

  6. [6]

    Bosma, Wieb and Cannon, John and Playoust, Catherine , title =. J. Symbolic Comput. , fjournal =. 1997 , number =. doi:10.1006/jsco.1996.0125 , url =

    work page doi:10.1006/jsco.1996.0125 1997

  7. [7]

    Decker, Wolfram and Greuel, Gert-Martin and Pfister, Gerhard and Sch\"onemann, Hans , year =

    work page

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    Gao, Guoxiong and Wang, Yutong and Jiang, Jiedong and Gao, Qi and Qin, Zihan and Xu, Tianyi and Dong, Bin , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  9. [9]

    doi:10.1038/s41586-025-09833-y , url =

    Nature , year =. doi:10.1038/s41586-025-09833-y , url =

    work page doi:10.1038/s41586-025-09833-y

  10. [10]

    Chervonyi, Yuri and Trinh, Trieu H. and Ol. Journal of Machine Learning Research , volume =. 2025 , url =

    work page 2025

  11. [11]

    DeepSeek-Prover: Advancing Theorem Proving in LLMs through Large-Scale Synthetic Data

    Xin, Huajian and others , title =. arXiv preprint arXiv:2405.14333 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Ren, Z. Z. and others , title =. arXiv preprint arXiv:2504.21801 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Kimina-Prover Preview: Towards Large Formal Reasoning Models with Reinforcement Learning

    Wang, Haiming and Unsal, Mert and Lin, Xiaohan and Baksys, Mantas and Liu, Junqi and others , title =. arXiv preprint arXiv:2504.11354 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving

    Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving , author =. arXiv preprint arXiv:2507.23726 , year =. 2507.23726 , archiveprefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    Shao, Zhihong and Luo, Yuxiang and Lu, Chengda and Ren, Z. Z. and others , title =. arXiv preprint arXiv:2511.22570 , year =

    work page arXiv

  16. [16]

    Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track , year =

    Balunovi. Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track , year =

    work page

  17. [17]

    Ma, Jicheng and others , journal =

    work page

  18. [18]

    and Motwani, Sumeet and Roggeveen, James V

    Wang, Erik Y. and Motwani, Sumeet and Roggeveen, James V. and others , journal =

    work page

  19. [19]

    2024 , note =

    Tsoukalas, George and Lee, Jasper and Jennings, John and Xin, Jimmy and Ding, Michelle and Jennings, Michael and Thakur, Amitayush and Chaudhuri, Swarat , booktitle =. 2024 , note =

    work page 2024

  20. [20]

    2026 , eprint =

    FATE: A Formal Benchmark Series for Frontier Algebra of Multiple Difficulty Levels , author =. 2026 , eprint =

    work page 2026

  21. [21]

    FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI

    Glazer, Elliot and others , title =. arXiv preprint arXiv:2411.04872 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track , year =

    Zhang, Jie and Petrui, Cezara and Nikoli. Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track , year =

    work page

  23. [23]

    2026 , howpublished =

    work page 2026

  24. [24]

    IMProofBench: Benchmarking AI on Research-Level Mathematical Proof Generation

    Schmitt, Johannes and B\'. arXiv preprint arXiv:2509.26076 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Shalyt, Michael and Elimelech, Rotem and Kaminer, Ido , journal =

    work page

  26. [26]

    AI for Math Workshop at the 42nd International Conference on Machine Learning (ICML) , year =

    Tang, Bintao and Yang, Xin and Wang, Yuhao and Qiu, Zixuan and Ji, Zimo and Jiang, Wenyuan , title =. AI for Math Workshop at the 42nd International Conference on Machine Learning (ICML) , year =

    work page

  27. [27]

    Nature , volume =

    Mathematical discoveries from program search with large language models , author =. Nature , volume =. 2024 , doi =

    work page 2024

  28. [28]

    2024 , eprint =

    Mining Math Conjectures from LLMs: A Pruning Approach , author =. 2024 , eprint =

    work page 2024

  29. [29]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    Novikov, Alexander and V. arXiv preprint arXiv:2506.13131 , year =. 2506.13131 , archiveprefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    International Conference on Learning Representations , volume =

    Tora: A tool-integrated reasoning agent for mathematical problem solving , author =. International Conference on Learning Representations , volume =

    work page

  31. [31]

    2024 , eprint =

    Das, Debrup and Banerjee, Debopriyo and Aditya, Somak and Kulkarni, Ashish , journal =. 2024 , eprint =

    work page 2024

  32. [32]

    ACM Computing Surveys , year =

    A Survey on Large Language Models for Mathematical Reasoning , author =. ACM Computing Surveys , year =. doi:10.1145/3786333 , note =

    work page doi:10.1145/3786333

  33. [33]

    International Conference on Learning Representations (ICLR) , year =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  34. [34]

    Context7: Up-to-date documentation for LLMs and AI code editors , year =

    work page

  35. [35]

    Claude Opus 4.7 System Card , year =

    work page

  36. [36]

    Introducing Claude Sonnet 5 , year =

    work page

  37. [37]

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models , year =. 2512.02556 , archiveprefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    DeepSeek V4 Technical Report , year =

    work page

  39. [39]

    arXiv preprint arXiv:2606.19348 , year =

    Deepseek-v4: Towards highly efficient million-token context intelligence , author =. arXiv preprint arXiv:2606.19348 , year =

    work page arXiv

  40. [40]

    GPT-5.5 System Card , year =

    work page

  41. [41]

    Gemini 3.5 Flash Model Card , year =

    work page

  42. [42]

    Gemini 3.1 Pro Model Card , year =

    work page

  43. [43]

    Qwen3.7: The Agent Frontier , year =

    work page

  44. [44]

    2026 , eprint =

    GLM-5: from Vibe Coding to Agentic Engineering , author =. 2026 , eprint =

    work page 2026

  45. [45]

    2024 , eprint =

    Hyperbolic knots with arbitrarily large torsion order in knot Floer homology , author =. 2024 , eprint =

    work page 2024

  46. [46]

    Kimi K2.7 Code , year =

    work page

  47. [47]

    MiniMax M3: Frontier Coding, 1M Context, Native Multimodality --- All in One Model , year =

    work page

  48. [48]

    PeerJ Computer Science , issn =

    SymPy: symbolic computing in Python , author =. PeerJ Computer Science , issn =

    work page

  49. [49]

    2022 , month =

    Chase, Harrison , title =. 2022 , month =

    work page 2022

  50. [50]

    Sakana Fugu Technical Report

    Sakana Fugu Technical Report , author =. arXiv preprint arXiv:2606.21228 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    2026 , eprint =

    Advancing Mathematics Research with AI-Driven Formal Proof Search , author =. 2026 , eprint =

    work page 2026

  52. [52]

    2026 , url =

    Planar Point Sets with Many Unit Distances , author =. 2026 , url =

    work page 2026

  53. [53]

    Hyperbolic knots with arbitrarily large torsion order in knot Floer homology

    Hyperbolic knots with arbitrarily large torsion order in knot Floer homology , author =. arXiv preprint arXiv:2412.20652 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    Pattern avoiding alternating involutions

    Pattern avoiding alternating involutions , author =. arXiv preprint arXiv:2206.13877 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  55. [55]

    Discrete & Computational Geometry , volume =

    Random meander model for links , author =. Discrete & Computational Geometry , volume =. 2024 , publisher =

    work page 2024

  56. [56]

    Indagationes Mathematicae , volume =

    Asymptotic analysis of Emden--Fowler type equation with an application to power flow models , author =. Indagationes Mathematicae , volume =. 2023 , publisher =

    work page 2023

  57. [57]

    Journal of Differential Equations , volume =

    Computer assisted proofs for transverse collision and near collision orbits in the restricted three body problem , author =. Journal of Differential Equations , volume =. 2023 , publisher =

    work page 2023

  58. [58]

    Journal of Pure and Applied Algebra , volume =

    Groups with small multiplicities of fields of values of irreducible characters , author =. Journal of Pure and Applied Algebra , volume =. 2023 , publisher =

    work page 2023

  59. [59]

    arXiv preprint arXiv:2412.02494 , year =

    On the hit problem for the polynomial algebra and the algebraic transfer , author =. arXiv preprint arXiv:2412.02494 , year =

    work page arXiv

  60. [60]

    The $(4,p)$-arithmetic hyperbolic lattices, $p\geq 2$, in three dimensions

    The (4, p) -arithmetic hyperbolic lattices, p 2 , in three dimensions , author =. arXiv preprint arXiv:2206.14174 , year =

    work page internal anchor Pith review Pith/arXiv arXiv