pith. sign in

arxiv: 2605.19338 · v1 · pith:R5CVQQSMnew · submitted 2026-05-19 · 💻 cs.MA · cs.AI· cs.CL

STAR-P\'olyaMath: Multi-Agent Reasoning under Persistent Meta-Strategic Supervision

Jiaao Wu , Xian Zhang , Hanzhang Liu , Sophia Zhang , Fan Yang , Yinpeng Dong This is my paper

Pith reviewed 2026-05-20 02:51 UTC · model grok-4.3

classification 💻 cs.MA cs.AIcs.CL
keywords multi-agent systemsmathematical reasoningmeta-strategic supervisionreasoner-verifier interactionstate machine orchestrationlong-horizon reasoningcompetition mathematics
0
0 comments X

The pith

STAR-PólyaMath uses a persistent Meta-Strategist to orchestrate multi-agent math reasoning and reach state-of-the-art results on eight top competition benchmarks.

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

The paper presents STAR-PólyaMath as a multi-agent framework designed to fix reliability problems in long-horizon mathematical reasoning, including hallucination buildup and memory loss. It structures the process as a state machine with nested challenge-step-replan loops run by a reasoning-free Python orchestrator that keeps control separate from inference. A central persistent Meta-Strategist tracks memory across attempts and supplies high-level guidance or directives to break out of unproductive paths. This combination produces perfect scores on AIME, Putnam, and HMMT benchmarks while posting its biggest gain on MathArena Apex 2025. Ablation tests indicate the performance edge stems from the orchestration structure itself rather than simply mixing different models.

Core claim

The authors claim that a multi-agent system built as an orchestrated state machine with nested loops and governed by a persistent Meta-Strategist that maintains cross-attempt memory and issues meta-level directives can systematically bound error propagation and deliver superior results on extended competition mathematics problems.

What carries the argument

The persistent Meta-Strategist, which maintains cross-attempt memory and supplies high-level strategic guidance or mandatory directives to steer the Reasoner-Verifier pairs out of unproductive loops.

If this is right

  • The framework produces perfect scores on AIME, Putnam, and HMMT 2025-2026 problems.
  • The largest reported margin appears on MathArena Apex 2025, where the system scores 93.75 percent against 80.21 percent for the strongest baseline.
  • Ablations confirm that removing key orchestration components or swapping backbones reduces results, pointing to the structure as the source of gains.
  • The design separates control flow from inference through a reasoning-free orchestrator that enables trace-back and re-planning.

Where Pith is reading between the lines

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

  • The same separation of meta-control from inference steps could be tested on non-math tasks that also require sustained multi-step planning.
  • Persistent memory across attempts might be adapted to domains where agents must learn from prior failed trajectories rather than resetting each time.
  • If the Meta-Strategist can issue directives without calling the underlying models, the approach may lower overall inference cost on long problems.

Load-bearing premise

The meta-level supervision and structured replanning loops can reliably prevent error accumulation without themselves introducing new inconsistencies or excessive overhead.

What would settle it

Measure performance on the same eight benchmarks with the Meta-Strategist component removed or disabled; a large drop relative to the full system would support the claim, while little or no change would indicate the orchestration is not the decisive factor.

Figures

Figures reproduced from arXiv: 2605.19338 by Fan Yang, Hanzhang Liu, Jiaao Wu, Sophia Zhang, Xian Zhang, Yinpeng Dong.

Figure 1
Figure 1. Figure 1: STAR-PólyaMath system workflow. STAR-PólyaMath advances each problem with four phases: exploration, planning and decomposition, step-wise execution with challenge loops, and solution generation. A Python orchestrator dispatches the LLM agents and decides advance, trace-back, re-plan, and abort transitions. The Reasoner probes the problem, proposes a step-wise plan, and executes each step with hierarchical … view at source ↗
Figure 2
Figure 2. Figure 2: Apex 2025 Problem 2 case study. (Left) The example single-pass GPT-5.5 baseline commits to the chain-of-pluses construction and the false universal bound k = 3/4. (Right) STAR￾PólyaMath’s Plan v1 falls into the same attractor; after three timeouts and trace-backs, the Meta￾Strategist’s cross-attempt memory diagnoses “the 3/4-bound is false, not unproved”, issues an APPROVE-REPLAN verdict with explicit forb… view at source ↗
Figure 3
Figure 3. Figure 3: Per-problem wall-clock distribution across the eight benchmarks, on a log scale. Each [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of verification tags across benchmarks, normalized to [PITH_FULL_IMAGE:figures/full_fig_p021_4.png] view at source ↗
read the original abstract

Frontier AI models and multi-agent systems have led to significant improvements in mathematical reasoning. However, for problems requiring extended, long-horizon reasoning, existing systems continue to suffer from fundamental reliability issues: hallucination accumulation, memory fragmentation, and imbalanced reasoning-tool trade-offs. In this paper, we introduce STAR-P\'olyaMath, a multi-agent framework that systematically addresses these challenges through meta-level supervision and structured Reasoner-Verifier interaction. STAR-P\'olyaMath is structured as an orchestrated state machine with nested challenge-step-replan loops, governed by a reasoning-free Python orchestrator that separates control from inference and bounds error propagation through trace-back and re-planning. Our key innovation is a persistent Meta-Strategist that maintains cross-attempt memory and exercises meta-level control by issuing high-level strategic guidance or mandatory directives, so the system can escape unproductive loops rather than stagnate or over-rely on tools. STAR-P\'olyaMath achieves state-of-the-art results on all eight top-tier competition benchmarks: AIME 2025-2026, MathArena Apex Shortlist, MathArena Apex 2025, Putnam 2025, IMO 2025, HMMT February 2026, and USAMO 2026. It obtains perfect scores on AIMEs, Putnam, and HMMT, and shows its largest margin on Apex 2025, scoring 93.75% compared with 80.21% by the strongest baseline GPT-5.5. Ablation studies show that the gains arise from the framework's orchestration rather than from model-level diversity since removing key components or substituting in mixed backbones consistently weakens performance. Code is available at https://github.com/Julius-Woo/STAR-PolyaMath.

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.

Referee Report

3 major / 3 minor

Summary. The paper introduces STAR-PólyaMath, a multi-agent framework for long-horizon mathematical reasoning. It features a reasoning-free Python orchestrator implementing nested challenge-step-replan loops, trace-back mechanisms, and a persistent Meta-Strategist that maintains cross-attempt memory and issues high-level strategic guidance or mandatory directives. The central empirical claim is state-of-the-art performance across eight competition benchmarks (AIME 2025-2026, MathArena Apex Shortlist, MathArena Apex 2025, Putnam 2025, IMO 2025, HMMT February 2026, USAMO 2026), including perfect scores on AIME, Putnam, and HMMT, with the largest margin on Apex 2025 (93.75% vs. 80.21% for GPT-5.5). Ablation studies are presented to attribute gains to the orchestration framework rather than model diversity.

Significance. If the performance claims and ablation attributions hold under detailed scrutiny, the work offers a concrete architecture for bounding error propagation and escaping unproductive loops in multi-agent reasoning systems. The separation of control logic into a reasoning-free orchestrator and the explicit cross-attempt memory mechanism address documented failure modes (hallucination accumulation, memory fragmentation) in a reproducible way; the public code release further strengthens the contribution by enabling direct replication and extension.

major comments (3)
  1. Ablation studies section: the statement that 'removing key components or substituting in mixed backbones consistently weakens performance' does not report whether the Persistent Meta-Strategist itself was ablated, held fixed at full strength, or replaced by a weaker model in the mixed-backbone runs. Because the central claim attributes gains to 'framework orchestration' rather than raw model capability, this omission is load-bearing; the skeptic note correctly identifies that any strategic directive must originate from an LLM component, so the current ablation design leaves open the possibility that observed improvements still trace to frontier-model strength rather than the nested meta-strategic structure.
  2. Methods / Experimental Setup: the manuscript does not specify the number of independent attempts, the exact prompting templates used by the Meta-Strategist, or the failure criteria that trigger re-planning versus mandatory directives. These details are required to evaluate whether the reported perfect scores on AIME/Putnam/HMMT reflect genuine escape from loops or simply higher per-attempt success rates of the underlying model.
  3. Results on MathArena Apex 2025: the 13.54-point margin over GPT-5.5 is the largest reported; however, without an error breakdown (e.g., percentage of problems solved only after Meta-Strategist intervention versus solved on first attempt), it is impossible to quantify how much of the margin is attributable to the persistent memory and meta-level control versus baseline model improvement.
minor comments (3)
  1. Abstract: the list of eight benchmarks appears to enumerate seven distinct contests (AIME 2025-2026 may be intended as two separate years); clarify the exact count and provide a table mapping each benchmark to its reported score and baseline.
  2. Notation: the term 'reasoning-free Python orchestrator' is used repeatedly but never formally defined; a short pseudocode block or state-machine diagram in §3 would remove ambiguity about which decisions are made outside any LLM call.
  3. Figure clarity: the state-machine diagram (if present) should explicitly label the trace-back and re-planning edges so readers can map them to the error-bounding claim in the abstract.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful and constructive comments on our manuscript. We address each of the major comments point by point below and describe the revisions we intend to make.

read point-by-point responses

  1. Referee: [—] Ablation studies section: the statement that 'removing key components or substituting in mixed backbones consistently weakens performance' does not report whether the Persistent Meta-Strategist itself was ablated, held fixed at full strength, or replaced by a weaker model in the mixed-backbone runs. Because the central claim attributes gains to 'framework orchestration' rather than raw model capability, this omission is load-bearing; the skeptic note correctly identifies that any strategic directive must originate from an LLM component, so the current ablation design leaves open the possibility that observed improvements still trace to frontier-model strength rather than the nested meta-strategic structure.

    Authors: The referee correctly notes that the current description does not specify the status of the Persistent Meta-Strategist in the ablations. We will revise the Ablation studies section to report in detail the configurations used for the Meta-Strategist across the ablation runs, including cases where it was ablated or replaced by weaker models. This revision will directly address the concern about whether improvements trace to the nested meta-strategic structure. revision: yes

  2. Referee: [—] Methods / Experimental Setup: the manuscript does not specify the number of independent attempts, the exact prompting templates used by the Meta-Strategist, or the failure criteria that trigger re-planning versus mandatory directives. These details are required to evaluate whether the reported perfect scores on AIME/Putnam/HMMT reflect genuine escape from loops or simply higher per-attempt success rates of the underlying model.

    Authors: We thank the referee for pointing out these omissions, which are important for reproducibility and for distinguishing the effects of the framework from baseline model performance. We will update the Methods / Experimental Setup section to include the number of independent attempts, the exact prompting templates for the Meta-Strategist, and the specific failure criteria that trigger re-planning versus mandatory directives. These details will be added to the main text or as supplementary material to allow readers to better evaluate the reported results. revision: yes

  3. Referee: [—] Results on MathArena Apex 2025: the 13.54-point margin over GPT-5.5 is the largest reported; however, without an error breakdown (e.g., percentage of problems solved only after Meta-Strategist intervention versus solved on first attempt), it is impossible to quantify how much of the margin is attributable to the persistent memory and meta-level control versus baseline model improvement.

    Authors: The referee makes a valid observation that an error breakdown would help quantify the contribution of the meta-level mechanisms to the performance margin on MathArena Apex 2025. We will incorporate such an analysis into the Results section, providing a breakdown of problems solved with and without Meta-Strategist intervention. This will offer a more precise attribution of the 13.54-point margin. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark claims rest on external comparisons

full rationale

The paper introduces a multi-agent framework (STAR-PólyaMath) with a reasoning-free Python orchestrator and persistent Meta-Strategist, then reports direct empirical results on eight external competition benchmarks (AIME, Putnam, IMO, etc.) plus ablation studies. No derivation chain, equations, fitted parameters renamed as predictions, or self-citation load-bearing steps appear in the manuscript. Central claims are falsifiable against public benchmarks and do not reduce to inputs by construction; ablations compare against mixed backbones and component removals without self-referential fitting. This is a standard empirical systems paper whose performance assertions are independent of any internal derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Review conducted on abstract only; no explicit free parameters, mathematical axioms, or invented physical entities are described. The Meta-Strategist is introduced as a new software component.

invented entities (1)
  • Persistent Meta-Strategist no independent evidence
    purpose: Maintains cross-attempt memory and issues high-level strategic guidance or directives to escape unproductive loops
    Presented as the key innovation that enables the system to avoid stagnation.

pith-pipeline@v0.9.0 · 5870 in / 1239 out tokens · 53846 ms · 2026-05-20T02:51:57.862384+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages · 3 internal anchors

  1. [1]

    Introducing GPT-5.2

    OpenAI. Introducing GPT-5.2. https://openai.com/index/introducing-gpt-5-2/ , 2025

    work page 2025

  2. [2]

    Gemini 3 pro

    Google DeepMind. Gemini 3 pro. https://deepmind.google/models/gemini/pro/, 2025

    work page 2025

  3. [3]

    The open proof corpus: A large-scale study of LLM-generated mathematical proofs.arXiv preprint arXiv:2506.21621, 2025

    Jasper Dekoninck, Ivo Petrov, Kristian Minchev, Mislav Balunovic, Martin Vechev, Miroslav Marinov, Maria Drencheva, Lyuba Konova, Milen Shumanov, Kaloyan Tsvetkov, et al. The open proof corpus: A large-scale study of LLM-generated mathematical proofs.arXiv preprint arXiv:2506.21621, 2025

    work page arXiv 2025

  4. [4]

    FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI

    Elliot Glazer, Ege Erdil, Tamay Besiroglu, Diego Chicharro, Evan Chen, Alex Gunning, Caroline Falkman Olsson, Jean-Stanislas Denain, Anson Ho, Emily de Oliveira Santos, et al. FrontierMath: A benchmark for evaluating advanced mathematical reasoning in AI.arXiv preprint arXiv:2411.04872, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems (NeurIPS), volume 35, pages 24824–24837, 2022

    work page 2022

  6. [6]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    The Lean 4 theorem prover and programming language

    Leonardo de Moura and Sebastian Ullrich. The Lean 4 theorem prover and programming language. InInternational Conference on Automated Deduction (CADE), pages 625–635. Springer, 2021

    work page 2021

  8. [8]

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

    Thomas Hubert, Remi Mehta, Laurent Sartran, et al. Olympiad-level formal mathematical reasoning with reinforcement learning.Nature, 2025. doi: 10.1038/s41586-025-09833-y

    work page doi:10.1038/s41586-025-09833-y 2025

  9. [9]

    doi:10.48550/arXiv.2512.17260 , url =

    Jiangjie Chen, Wenxiang Chen, Jiacheng Du, Jinyi Hu, Zhicheng Jiang, Allan Jie, Xiaoran Jin, Xing Jin, Cheng Li, Zheng Yuan, et al. Seed-Prover 1.5: Mastering undergraduate-level theorem proving via learning from experience.arXiv preprint arXiv:2512.17260, 2025

    work page arXiv 2025

  10. [10]

    ToRA: A tool-integrated reasoning agent for mathematical problem solving

    Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. ToRA: A tool-integrated reasoning agent for mathematical problem solving. InInterna- tional Conference on Learning Representations (ICLR), 2024. 10

    work page 2024

  11. [11]

    PAL: Program-aided language models

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. PAL: Program-aided language models. InInternational Conference on Machine Learning (ICML), pages 10764–10799, 2023

    work page 2023

  12. [12]

    Large language models cannot self-correct reasoning yet

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xiny- ing Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In International Conference on Learning Representations (ICLR), 2024

    work page 2024

  13. [13]

    MACM: Utilizing a multi-agent system for condition mining in solving complex mathematical problems

    Bin Lei, Yi Zhang, Shan Zuo, Ali Payani, and Caiwen Ding. MACM: Utilizing a multi-agent system for condition mining in solving complex mathematical problems. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  14. [14]

    Yichen Huang and Lin F. Yang. Winning gold at IMO 2025 with a model-agnostic verification- and-refinement pipeline.arXiv preprint arXiv:2507.15855, 2025

    work page arXiv 2025

  15. [15]

    Bowman, Trevor Darrell, and Ethan Perez

    Akbir Khan, John Hughes, Dan Valentine, Laura Ruis, Kshitij Sachan, Ansh Radhakrishnan, Edward Grefenstette, Samuel R. Bowman, Trevor Darrell, and Ethan Perez. Debating with more persuasive LLMs leads to more truthful answers. InInternational Conference on Machine Learning (ICML), 2024

    work page 2024

  16. [16]

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. AutoGen: Enabling next-gen LLM applications via multi-agent conversation.arXiv preprint arXiv:2308.08155, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [17]

    Hilbert: Recursively building formal proofs with informal reasoning.arXiv preprint arXiv:2509.22819, 2025

    Sumanth Varambally, Thomas V oice, Yanchao Sun, Zhifeng Chen, Rose Yu, and Ke Ye. Hilbert: Recursively building formal proofs with informal reasoning.arXiv preprint arXiv:2509.22819, 2025

    work page arXiv 2025

  18. [18]

    Toolorchestra: Elevating intelligence via efficient model and tool orchestration

    Hongjin Su, Shizhe Diao, Ximing Lu, Mingjie Liu, Jiacheng Xu, Xin Dong, Yonggan Fu, Peter Belcak, Hanrong Ye, et al. ToolOrchestra: Elevating intelligence via efficient model and tool orchestration.arXiv preprint arXiv:2511.21689, 2025

    work page arXiv 2025

  19. [19]

    Zhihong Shao, Yuxiang Luo, Chengda Lu, Z. Z. Ren, Jiewen Hu, Tian Ye, Zhibin Gou, Shirong Ma, and Xiaokang Zhang. DeepSeekMath-V2: Towards self-verifiable mathematical reasoning. arXiv preprint arXiv:2511.22570, 2025

    work page arXiv 2025

  20. [20]

    Brains vs

    Hamed Mahdavi, Alireza Hashemi, Majid Daliri, Pegah Mohammadipour, Alireza Farhadi, Samira Malek, Yekta Yazdanifard, Amir Khasahmadi, and Vasant Honavar. Brains vs. bytes: Evaluating LLM proficiency in olympiad mathematics.arXiv preprint arXiv:2504.01995, 2025

    work page arXiv 2025

  21. [21]

    To code or not to code? adaptive tool integration for math language models

    Jiaheng Wang et al. To code or not to code? adaptive tool integration for math language models. arXiv preprint, 2025

    work page 2025

  22. [22]

    Diverse inference and verification for advanced reasoning.arXiv preprint arXiv:2502.09955, 2025

    Iddo Drori, Gaston Longhitano, Mao Mao, Seunghwan Hyun, Yuke Zhang, Sungjun Park, Zachary Meeks, Xin-Yu Zhang, and Ben Segev. Diverse inference and verification for advanced reasoning.arXiv preprint arXiv:2502.09955, 2025

    work page arXiv 2025

  23. [23]

    Princeton University Press, 1945

    George Pólya.How to Solve It: A New Aspect of Mathematical Method. Princeton University Press, 1945

    work page 1945

  24. [24]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. InInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  25. [25]

    Solving challenging math word problems using GPT-4 code interpreter with code-based self-verification

    Aojun Zhou, Ke Wang, Zimu Lu, Weikang Shi, Sichun Luo, Zipeng Qin, Shaoqing Lu, Anya Jia, Linqi Song, Mingjie Zhan, and Hongsheng Li. Solving challenging math word problems using GPT-4 code interpreter with code-based self-verification. InInternational Conference on Learning Representations (ICLR), 2024

    work page 2024

  26. [26]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. InInternational Conference on Machine Learning (ICML), 2024. 11

    work page 2024

  27. [27]

    CAMEL: Communicative agents for "mind" exploration of large language model society

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. CAMEL: Communicative agents for "mind" exploration of large language model society. InAdvances in Neural Information Processing Systems (NeurIPS), volume 36, 2023

    work page 2023

  28. [28]

    Sumeet Ramesh Motwani, Chandler Smith, Rocktim Jyoti Das, Rafael Rafailov, Ivan Laptev, Philip H. S. Torr, Fabio Pizzati, Ronald Clark, and Christian Schroeder de Witt. MALT: Improving reasoning with multi-agent LLM training. InInternational Conference on Machine Learning (ICML), 2025

    work page 2025

  29. [29]

    ReMA: Learning to meta-think for LLMs with multi-agent reinforcement learning

    Ziyu Wan, Yunxiang Li, Xiaoyu Wen, Yan Song, Hanjing Wang, Linyi Yang, Mark Schmidt, Jun Wang, Weinan Zhang, Shuyue Hu, and Ying Wen. ReMA: Learning to meta-think for LLMs with multi-agent reinforcement learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    work page 2025

  30. [30]

    Self- refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self- refine: Iterative refinement with self-feedback. InAdvances in Neural Information Processing Sy...

    work page 2023

  31. [31]

    Proof or bluff? evaluating LLMs on 2025 USA math olympiad.arXiv preprint arXiv:2503.21934, 2025

    Ivo Petrov, Jasper Dekoninck, Lyuben Baltadzhiev, Maria Drencheva, Kristian Minchev, Mislav Balunovi´c, Nikola Jovanovi´c, and Martin Vechev. Proof or bluff? evaluating LLMs on 2025 USA math olympiad.arXiv preprint arXiv:2503.21934, 2025

    work page arXiv 2025

  32. [32]

    Trinh, Yuhuai Wu, Quoc V

    Trieu H. Trinh, Yuhuai Wu, Quoc V . Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations.Nature, 625:476–482, 2024

    work page 2024

  33. [33]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large language models.Nature, 625:468–475, 2024

    work page 2024

  34. [34]

    Agentic reasoning: A streamlined framework for enhancing LLM reasoning with agentic tools

    Jingyuan Wu et al. Agentic reasoning: A streamlined framework for enhancing LLM reasoning with agentic tools. InProceedings of ACL, 2025

    work page 2025

  35. [35]

    Math- Arena: Evaluating LLMs on uncontaminated math competitions, 2025

    Mislav Balunovi´c, Jasper Dekoninck, Ivo Petrov, Nikola Jovanovi´c, and Martin Vechev. Math- Arena: Evaluating LLMs on uncontaminated math competitions, 2025

    work page 2025

  36. [36]

    MathArena Apex: Uncon- quered final-answer problems, 2025

    Jasper Dekoninck, Nikola Jovanovi´c, Ivo Petrov, and Martin Vechev. MathArena Apex: Uncon- quered final-answer problems, 2025. URLhttps://matharena.ai/apex/. A System Configuration and Budgets Runtime and Loop Bounds.Table 4 summarizes the hard limits governing the STAR-PólyaMath orchestrator. All timeouts apply to LLM inference and verification code exec...

    work page 2025

  37. [37]

    Plan Creation: decompose the problem into 3-10 numbered steps

    work page

  38. [38]

    Step Execution: execute one step at a time with rigorous reasoning

    work page

  39. [39]

    Verification Tagging: tag every nontrivial claim with one of [verified] / [easy-verify] / [hard-verify]

    work page

  40. [40]

    Challenge Response: address the Verifier’s concerns with evidence

    work page

  41. [41]

    trace back

    Code Execution: run code yourself to verify computational claims; do not propose code with [easy-verify] when you can run it. ## Verification Tags Load the ‘verification-tag-protocol‘ skill for full definitions. - [verified] - You actually ran code; report the real output. - [easy-verify] - Use only when you cannot run code yourself. - [hard-verify] - Log...

    work page 2025

  42. [42]

    Fix the exact combinatorial model

    work page

  43. [43]

    Split the counts into inside and outside contributions

    work page

  44. [44]

    Find the correct candidate constant from fresh evidence

    work page

  45. [45]

    Prove the universal lower bound

    work page

  46. [46]

    Make the discharging proof independent of shape pathologies

    work page

  47. [47]

    Construct polygons approaching equality

    work page

  48. [48]

    Conclude sharpness and the largest value. Step 3 (new candidate).Re-running candidate search on a denser 4×4 motif (rows#.#.,####, #.#.,....) tiledn×nproduced a1 = 9n, a 2 = 4n 2 −n, a 3 = 8n 2 −n, and forn= 50the ratio is208/399≈0.5213— already well below3/4atn= 5. Step 4 (cap-map proof).The verified lemmas establishing a3 ≤a 1 + 2a2 are recorded in the ...

    work page 2025

  49. [49]

    Recast the grid condition exactly. 23

    work page

  50. [50]

    Replace tilings by a rectilinear-geometry invariant

    work page

  51. [51]

    Specialize good chords to permutation geometry

    work page

  52. [52]

    Translate chord selection into a bipartite matching problem

    work page

  53. [53]

    Independently search for the extremal construction

    work page

  54. [54]

    Prove the construction’s upper bound geometrically

    work page

  55. [55]

    Prove the matching lower bound for every permutation

    work page

  56. [56]

    Confirmed Failures

    Assemble the final equality. Reformulation phase (Steps 1–3).Step 1 establishes that any valid configuration has uncovered set Uπ ={(i, π(i)) : 1≤i≤2025} for some permutation π∈S 2025, and conversely; the problem reduces to minπ T(π) , where T(π) is the minimum number of rectangles partitioning the complement ofU π. Step 2 entered a multi-round debate. Th...

    work page 2025