acceptodds
Under review as a conference paper at ICLR 2027

Context-Orchestrated Research Math Agents

Abstract

Long-horizon mathematical reasoning fails less often because a model cannot produce a valid next step than because an agent fails to maintain and expose the right semantic state across many iterations. Left unmanaged, this produces research-level proofs that are locally convincing yet globally incomplete: a key lemma unproved, an assumption unchecked, a citation unsupported, or a computational claim unverified. We present **Research Math Agents (RMA)**, an agentic framework for long-horizon proof development built around a persistent, typed research store and an orchestrator that compiles operation-specific context from that store. The **Research Context Orchestrator** is the central state-management layer between the persistent research store and each locally scoped proof operation: it retrieves task-relevant artifacts, compiles them into a bounded context, invokes the appropriate operation, and writes the resulting proof edits, issue updates, literature notes, plans, or evaluations back to the store. This process is designed to keep proof revisions, unresolved issues, prior attempts, literature, and evaluations available across rounds while exposing only task-relevant state to each local operation. We evaluate RMA across complementary research-level settings using independent expert evaluation, blind mathematician review, LLM-based benchmark evaluation, and Lean 4 kernel verification. RMA achieves a **42.5% solve rate** on the independently evaluated SOOHAK Challenge Hard set, obtains **8 of 10 correct solutions** on First Proof B1 and **8 of 10 passing solutions** on B2 under human-expert evaluation, and verifies **213 of 300** sampled Research Solved targets in Formal Conjectures with the Lean 4 kernel. These results support structured state and context orchestration as useful components for long-horizon proof development, while comparisons against independently developed systems remain subject to differences in inference budgets, backbones, and evaluation protocols. Project page: https://anonymous1776research.github.io/RMA-ICLR2027/

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, long-horizon reasoning, mathematical reasoning, context orchestration, persistent memory, iterative refinement, automated theorem proving

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_Tu3nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new agentic framework for research-level math problem-solving, called RMA. This workflow addresses the inconsistency issue in long-horizon math reasoning with a typed research store and a context orchestrator. In the following experiments with human-expert evaluation, this workflow constantly exhibits superior performance over several baselines.

Critical strengths
  • The superior performance of RMA is supported by many experiments on frontier math research. The authors also reported the cost and scalability of each experiments, providing more information on the efficiency of RMA.
  • The inconsistency of agents in the long-horizon math research is an important issue in related research. This work is actually trying to address some realistic and meaningful drawbacks.
  • The authors have done substantial experiments on research-level math problems. The human-expert evaluation do require a large amount of work.
Critical weaknesses
  • The generalizability might be limited by the complex workflow design of RMA. Its implications for subsequent work would therefore also be limited.
  • Some experiments and claims have minor reliability issues (See Question 2 below).
Questions for authors
  • Among all parts in the agent workflow design, which one do you believe has the highest contribution to the final performance?
  • The performance in the component ablation study (Figure 2, c) seems too good to be true. I am wondering about the randomness in these experiments. Since These experiments used a complex workflow with LLM judge and the dataset only consists 30 problems, I believe there should be notable variance in the score estimation. How do you get a monotone performance gain in the ablation study? The claim that all components be effective is not very convincing.
  • Have you considered comparing with some general-purpose agents such as OpenAI Codex or Claude Code on these long-horizon math problems? These are, as far as I know, famous tools that a mathematician will use in their daily work as well. So I believe this comparison would be very informative.

Minor Issues or typos:

  • The right subfigure of Table 2 is not correctly sorted by pass rate as declared in the table.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Tu3n (part 1/2)

We thank you for recognizing the importance of long-horizon consistency, the breadth of the research-level experiments, and the effort involved in expert evaluation.

W1. Generalizability of the workflow.

RMA targets mathematical research, while its implications extend to how agents organize state and coordinate extended work.

  1. Mathematical research is itself a substantial and demanding application domain. Developing research-level proofs requires tracking assumptions, unresolved lemmas, failed attempts, and dependencies across an evolving argument. RMA's operations address these requirements. Our evaluation spans informal proof development and Lean-verified theorem proving, with mathematician assessment and kernel verification. Improving agents' ability to develop research-level mathematics is a meaningful objective in its own right.

  2. Recent work shows that related design needs extend beyond mathematics. Anthropic's 2026 study of long-running application development [1] coordinates planner, generator, and evaluator agents through structured files and explicit acceptance criteria. YouRA [2], released in October 2026, maintains hypotheses, execution evidence, and failure histories to control recovery and review in autonomous research. These systems illustrate a broader challenge: agents must preserve unfinished work and accumulated evidence while deciding what to do next. RMA examines this challenge through proof-specific state and issue-directed orchestration.

  3. Subsequent systems can adapt the state–operation interface to their own tasks. In software engineering, unresolved proof obligations could become failing tests or unmet requirements, with attempted patches and test results retained for later operations. In experimental research, the records could represent hypotheses, unsuccessful experiments, and outstanding validation checks. The operations and validation tools would change, while the store and orchestrator could preserve dependencies, retrieve relevant prior work, and record outcomes. Our ablations provide concrete evidence about the effects of persistent state, context construction, and scheduling that can inform these adaptations.

We have revised the related works section to reflect this discussion and we have revised the conclusion section to discuss future directions.

W2. Randomness and the ablation results.

The monotone appearance comes from sorting the measured scores for display. Each component is tested separately against the full system; the bars do not represent successive component removals or additions. We have clarified this in the Figure 2(c) caption and the “Component ablations” paragraph in Section 4.2.

  1. Each bar is a separate experiment. For example, the literature ablation retains meetings and coordinated revision, while the meetings-and-revision ablation retains literature grounding. We evaluate these configurations individually and then arrange their scores in ascending order. The visual ordering therefore does not imply a monotone sequence of experimental improvements.

  2. We quantify run-to-run variation through repeated experiments, summarized in the table below. Six central configurations are evaluated over three runs on the same 30 problems. Full RMA scores 8.53 ± 0.15, severity-only ordering 7.23 ± 0.15, context dumping 3.47 ± 0.21, and the stateless store 2.63 ± 0.25. Here, ± denotes the standard deviation across run-level mean scores. Each proof is anonymized and assessed by the same three-model judge panel.

  3. The repeated results support the larger observed effects. Sorting the single-run bars does not establish significance between nearby configurations. The central state-management finding also receives separate blinded mathematician evaluation: full RMA achieves 25/30 passes versus 16/30 for summary memory, with a two-sided exact McNemar p = 0.049.

Table. Repeated key architectural comparisons

ConfigurationMean continuous score ± standard deviation
Full RMA8.53 ± 0.15
Severity-only issue ordering7.23 ± 0.15
Summary-memory RMA6.63 ± 0.15
GIR with RMA's output-selection rule5.50 ± 0.10
Context dumping3.47 ± 0.21
Stateless store2.63 ± 0.25
AuthorsOpenReview ↗

Response to Reviewer Tu3n (part 2/2)

Q1. Which part contributes most to final performance?

Persistent state and bounded context construction produce the largest degradations when weakened. In the plotted ablation, replacing the store with a stateless variant reduces the score from 8.5 to 2.6. Replacing bounded context construction with context dumping reduces it to 3.4. The repeated means are 2.63 ± 0.25 and 3.47 ± 0.21, compared with 8.53 ± 0.15 for full RMA. The more controlled summary-memory comparison strengthens the conclusion about persistent state. It retains the full workflow and yields 6.63 ± 0.15, with 16/30 human passes compared with 25/30 for full RMA. The context-dumping result requires a narrower interpretation. Dumping also produces oversized or failed calls and fewer completed operations. Its degradation therefore reflects a joint information-and-resource effect, rather than a pure effect of irrelevant context. These results place persistent research state and controlled exposure of that state among the most consequential mechanisms in the tested architecture.

Q2. How are the ablations scored, and why do the results look consistently ordered?

The continuous score measures proof correctness and completeness. Candidate proofs are anonymized and scored from 0 to 10 by Opus 4.8, GPT-5.2, and Gemini 3.1 Pro. Their scores are averaged for each proof and then across problems. Each proof is assessed once by each judge by default. Repeated experimental results come from repeated system runs, rather than repeatedly judging one fixed proof. Same set of problems are used in all ablation studies.

Figure 2(d) contains three separate comparisons:

  • Stateless memory, previous-round-only memory, and full persistence.
  • Context dumping, naive truncation, and operation-specific compilation.
  • FIFO, severity-only, and severity-plus-dependency issue ordering.

The intermediate settings are included to test milder alternatives, rather than relying only on extreme component removal. Some intermediate settings are single-run measurements, so their exact ordering should not be interpreted as a statistical ranking.

Blinded mathematician evaluation gives 25/30 passes for full RMA and 16/30 for summary memory. On 13 problems, only RMA passes; on 4, only summary memory passes (two-sided exact McNemar p = 0.049).

The result order question is answered in W2.

Q3. Comparison with Codex or Claude Code.

Most of our API budget is allocated to Claude models, so our controlled workflow comparisons primarily use Claude backbones. We use GPT-5.2 and Gemini 3.1 Pro alongside Opus 4.8 for evaluation under a small budget.

We compare different agent frameworks under the same backbone across several experiments: RMA and GIR use Opus 4.8, identical tools, and $185 per problem on First Proof B2; RMA and agentic baselines use Opus 4.8, identical tools, and same budget per problem on held-out ResearchMath. The summary-memory comparison also retains RMA's backbone and workflow while changing state representation and retrieval.

We have additionally included a three-backbone comparison in Section 4.2. For Opus 4.6, Opus 4.8, and Fable 5.1, we compare RMA with GIR equipped with RMA's output-selection rule on the same 30 held-out problems, using identical tools and the cost cap. RMA leads by 2.32, 3.03, and 3.08 points, respectively. These comparisons prioritize identifying the effect of workflow organization under shared conditions; a direct Codex or Claude Code comparison would additionally evaluate those systems' own orchestration and context management.

Minor comment. Sorting in Table 2.

Thank you for pointing this out. We have reordered the Formal Conjectures panel in Table 2 by verified count, from highest to lowest.


Overall, we have clarified that Figure 2(c) reports separate component ablations sorted by score, provided repeated-run results, and added a three-backbone comparison to Section 4.2. The repeated experiments and blinded human evaluation support the value of structured research state and bounded context construction. We also expanded Related Work and the conclusion to explain the design's relevance beyond mathematics, and corrected the ordering in Table 2. If these results and clarifications address your concerns, we would greatly appreciate it if you would consider raising your score.

References:

[1] Rajasekaran, P. (2026). Harness design for long-running application development. Anthropic, March 24. https://www.anthropic.com/research/harness-design-long-running-apps

[2] Woo, Y., Lee, W., and Huang, J.-X. (2026). YouRA: A Persistent-State Architecture for Evidence-Traceable Autonomous Research Agents. arXiv:2610.01097. https://arxiv.org/abs/2610.01097

I have read your rebuttal and I believe my major concern in soundness is resolved. The deviation of scores in the repeated experiments reported is very informative. I think the contribution of this paper to the ML or AI4Math community is still incremental. Given the limited significance I will not raise my score to 4. Nevertheless, I would like to see this paper being accepted for their hard work and concrete experimental results.

We sincerely thank the reviewer for their thoughtful feedback and for acknowledging that our rebuttal has resolved the major soundness concern. We are also glad that the additional repeated experiments and reported score variations were informative.

Regarding significance, we believe our work provides valuable insights into the practical challenges and reliability of AI for mathematical reasoning, supported by systematic experiments and concrete empirical evidence. We thank the reviewer again for their constructive feedback and encouraging comments.

Reviewer_qVnoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper presents Research Math Agents (RMA), an agentic framework for long-horizon proof development built around a persistent, typed research store and a Research Context Orchestrator, which retrieves task-relevant artifacts, compiles them into a bounded context for each locally scoped proof operation, and writes the results back to the store. Across complementary research-level settings, including independent expert evaluation, blind mathematician review, and Lean 4 kernel verification, RMA shows strong performance, which supports structured state and context orchestration as useful components for long-horizon proof development.

Critical strengths
  1. The paper focuses on state management in long-horizon proving, which is a key reason why research-level proofs end up incomplete. The structured state design improves performance clearly over simply summary memory, and superior to many other systems.

  2. The paper has a thorough evaluation. The experiments span four benchmarks with complementary protocols (third-party, blinded expert, LLM judge, and Lean verification), along with extensive ablations and detailed leakage controls which gives sufficient information.

Critical weaknesses

On B2, the paper reports realized costs and a GIR control under the same budget, which is convincing. The other headline comparisons lack this support. On SOOHAK Challenge Hard and Formal Conjectures, RMA's realized spend is not reported. Meanwhile on Formal Conjectures, the direct Claude Opus 4.8 baseline lacks RMA's issue-repair loop, so it receives no Lean feedback, and its number of attempts is not reported. The gaps on these benchmarks (42.5% vs. 12.5%, 213/300 vs. 121/300) may therefore partly reflect additional inference rather than the architecture.

Questions for authors
  1. The abstract opens with the claim that long-horizon failures stem more from state management than from producing valid next steps. Is this based on an analysis of failure cases, or is it a working hypothesis?

  2. The ethics statement recommends using RMA with a human in the loop. How can a mathematician inspect and edit the research store during a run, for example by adding or closing issues?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qVno

We thank you for recognizing the importance of structured state, the breadth of the evaluation, and the value of the budget-matched GIR control.

W1. Additional inference versus architectural improvement.

We have supplemented the resource reporting and added controlled comparisons that directly test the architectural benefit.

  1. Formal Conjectures now includes realized expenditure and the common budget constraint. The implementation appendix reports RMA's total API cost of \2,738 on the 300 targets. Section 4.1 specifies the same \10-per-target ceiling across all attempts, the same target IDs, and pinned Lean environments. The systems retain their respective inference procedures, so a common monetary ceiling does not imply identical call counts or repair attempts. The direct Claude baseline provides a same-backbone reference without Lean feedback during proof generation, making it a weaker baseline than an interactive proving system. However, our Formal Conjectures evaluation also includes stronger specialized systems, including Seed-Prover 1.5 [1] and Hilbert [2], using their native inference procedures. On the same 300 targets, RMA verifies 213, compared with 193 for Seed-Prover and 164 for Hilbert. Thus, the evaluation compares RMA with both its standalone backbone and specialized theorem-proving systems.

  2. SOOHAK is independently executed under the organizers' protocol. RMA and the reported baselines receive the same \$4,000 budget ceiling, and the organizers control execution and grading over three seeded repetitions. This ceiling is distinct from realized expenditure. The reported gap measures end-to-end performance under that benchmark protocol.

  3. The controlled experiments provide the evidence that workflow organization matters under a shared resource allowance. On First Proof B2, RMA and GIR use the same backbone, tools, and \185−per−problembudget,achieving8/10versus4/10passes.Thebackbone−sensitivityexperimentnowincludedinSection4.2furthercomparesRMAwithGIRusingRMA′soutput−selectionruleundera185-per-problem budget, achieving 8/10 versus 4/10 passes. The backbone-sensitivity experiment now included in Section 4.2 further compares RMA with GIR using RMA's output-selection rule under a \\100-per-problem cap. RMA leads by 2.32–3.08 points across three backbones. These comparisons show that the measured benefit persists when the backbone, available tools, and monetary allowance are controlled.


Q1. Is the opening claim about failure causes established empirically?

This was a working hypothesis motivating the design, rather than a conclusion from a comparative analysis of failure causes. We have revised the abstract's opening sentence to make this clear:

“Long-horizon mathematical reasoning requires both valid local reasoning and persistent access to the evolving research state across iterations.”

Our controlled ablations examine how research-state representation and context construction affect proof quality: replacing typed records and record-level retrieval with summary memory reduces blinded human passes from 25/30 to 16/30, while replacing bounded context orchestration with full-state dumping significantly reduces the mean proof score from 8.53 to 3.47.

Q2. How can a mathematician inspect and edit the store during a run?

Good question. The reported implementation supports inspection through separately recorded research artifacts. A mathematician can examine proof revisions, issue records, literature notes, meeting outputs, concepts, strategic insights, and evaluations. Round-level telemetry also records completeness scores, proved-leaf fractions, and opened and closed issues. These records make it possible to inspect what the system has established, what it is attempting to repair, and what sources it has used.

We thank the reviewer for suggesting live editing. We have implemented this feature in our codebase and plan to include it in the public code release upon publication. The experiments reported in the paper use automated runs without manual edits.


Overall, we have expanded resource reporting, clarified the Formal Conjectures baseline settings, and added backbone-sensitivity results showing that RMA's advantage over GIR persists across three models under shared budget constraints. We also revised the abstract to remove the unsupported ranking of failure causes and explained how mathematicians can inspect the recorded research state. Together, the controlled comparisons and benchmark evaluations provide clearer evidence of RMA's practical value. If these results and clarifications address your concerns, we would greatly appreciate it if you would consider raising your score.

References:

[1] Chen, J., et al. (2025). Seed-Prover 1.5: Mastering Undergraduate-Level Theorem Proving via Learning from Experience. https://arxiv.org/abs/2512.17260

[2] Varambally, S., et al. (2026). Hilbert: Recursively Building Formal Proofs with Informal Reasoning. ICLR 2026. https://arxiv.org/abs/2509.22819

Reviewer_QjBGOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors propose Research Math Agents (RMA), a framework designed for long-horizon mathematical proof development. The framework consists of two core components: a persistent Research Store and a Research Context Orchestrator. The Research Store maintains structured information across multiple reasoning rounds, including proof revisions, unresolved issues, previous attempts, relevant literature, and evaluation results. The Research Context Orchestrator retrieves task-relevant information from the store, constructs bounded, operation-specific contexts for different research operations, such as critique, solving, revision, and evaluation, and writes the resulting updates back to the store. This design aims to improve the continuity and completeness of long-horizon mathematical reasoning. The authors evaluate their framework across several research-level mathematical benchmarks, including SOOHAK, First Proof, ResearchMath, and Formal Conjectures, using multiple evaluation protocols such as human expert review, LLM-based judging, and Lean 4 verification. They also conduct extensive ablation studies to investigate the contributions of different architectural components.

Critical strengths
  1. The paper is aimed to solve a critical problem in long-horizon proof development that AI-generated proofs are often locally convincing but globally incomplete.
  2. In the manuscript, the authors employ multiple complementary evaluation methods which is convincing.
Critical weaknesses
  1. One of the paper’s core innovations is the Long-Term Research Store. However, similar concepts have already been explored in several prior works, including Danus, ProofCouncil, and MechMath Agent Team, through mechanisms such as fact graphs and long-term memory. The paper does not clearly articulate how its approach differs from these methods, particularly in terms of maintaining reasoning states.
  2. The authors explain why the chosen operations are reasonable, but do not establish their necessity. And some operations may overlap in their roles such as solver and revision.
  3. The paper emphasizes preventing incomplete proofs by explicitly tracking unresolved proof obligations. However, Section 3.2 states that an issue is marked as resolved once the solver’s patch or rewrite has been successfully written back, without immediately verifying whether the corresponding mathematical gap has actually been closed. This may affect the evaluation of the final result.
Questions for authors
  1. Could the authors explore whether alternative combinations of operations could outperform the proposed design, and whether a simpler configuration could achieve comparable performance?
  2. Some details are not consistent. For example, Table 2 reports 7, 3, and 3 passing solutions for ProofCouncil, UCLA Moonshot, and ChatGPT 5.5 Pro, respectively, whereas Figure 2(a) appears to report 6, 5, and 5. Authors should check these inconsistence and other details in the paper.
  3. Could the authors more clearly demonstrate the practical effectiveness of RMA? For example, compared with the summary-based baseline, does RMA better preserve proof obligations and records of failed attempts, and enable more successful reuse of previously established results on certain problems? A few case studies would be particularly helpful for understanding how these mechanisms work.
4 comments
AuthorsOpenReview ↗

Response to Reviewer QjBG (part 1/2)

We thank you for recognizing the importance of globally complete proofs and the value of the complementary evaluation methods.

W1. Distinction from other research-state architectures.

Danus [1], ProofCouncil [2], and MechMath Agent Team [3] are concurrent works that share our interest in maintaining state for mathematical research. We have strengthened both the experimental comparison and the discussion of their relationship to RMA.

  1. We revised Related Work (Section 2) to explain the methodological differences explicitly. Danus [1] maintains verified claims in a fact graph alongside exploratory memory. ProofCouncil [2] maintains proof files and research notes within an author–critic workflow. MechMath Agent Team [3] combines task ledgers, file-based handoffs, and a knowledge base. RMA differs from these concurrent systems in how persistent state governs proof development: (a) Research state is organized around unresolved proof obligations. Proof revisions are linked to issues, prior attempts, supporting evidence, meeting plans, and evaluations. This preserves both established arguments and the reasons further work is needed. (b) Orchestration explicitly connects state to action. The Research Context Orchestrator prioritizes issues by severity and dependency impact, retrieves operation-specific evidence, and coordinates critic, solver, literature, meeting, revision, and evaluation operations. (c) Revision is coupled with continued assessment. Anchored local repairs constrain changes to selected proof spans, meeting plans guide broader revisions, and subsequent critic checks revisit the updated argument. Research outcomes are written back into the store to inform later rounds.
  2. We added a direct comparison with Danus [1] under the same backbone and budget (Section 4.2). We evaluate its released implementation on the same 30 held-out ResearchMath problems, using Claude Opus 4.8, identical tools, and a $100-per-problem API budget. Danus retains its own prompts, fact-graph memory, and orchestration workflow. Both systems use the same three-model judge panel and rubric, with separate blinded mathematician evaluation. RMA scores 8.53 ± 0.15 versus 8.04 ± 0.18 for Danus and achieves 25/30 versus 20/30 human passes. We observe an interesting phenomenon: more rounds do not always produce better proofs. Later Danus revisions sometimes weakened earlier arguments, while RMA better preserved prior progress. RMA anchored local repairs, persistent issue records, and repeated critic checks help focus further edits on unresolved gaps while preserving established reasoning.
  3. ProofCouncil is already included in our First Proof B2 comparison (Section 4.1). The rebuttal period limits the scale of further experiments (especially when humans are involved), so we prioritize a focused comparison on the existing held-out set rather than attempting large-scale evaluations of all three systems.

W2. Necessity and overlap of the operations.

Section 4.2 provides ablation evidence for the operations, and we have clarified the distinction between solver and revision in Section 3.2.

  1. The ablations measure their practical contribution. On the same 30 held-out ResearchMath problems under a $100-per-problem cap, removing meetings and coordinated revision while retaining local solver repairs reduces the score from 8.5 to 6.4. Disabling literature grounding gives 4.2. Removing individual critic analyses yields scores between 5.4 and 6.4, while removing evaluator feedback gives 7.5. These results support the usefulness of the tested operations within RMA.

  2. Solver and revision serve different scopes of proof development. The solver receives one selected issue and returns an anchored patch, preserving text outside the edited span. Revision receives a meeting-generated action plan and can coordinate changes across the proof, such as reorganizing an argument or changing its strategy. The solver repairs a specific gap; revision addresses changes that require several parts of the argument to be reconsidered together. Section 3.2 and the operation prompts in the appendix describe these interfaces.

W3. Meaning of “resolved” and evaluation of final proofs.

We have clarified Section 3.2 to distinguish an applied repair from a verified mathematical resolution.

  1. “Resolved” records that a repair has been applied. It does not certify that the gap is closed. There is no immediate issue-specific verification. The next critic round examines the updated proof and records remaining or reintroduced defects as new issues.

  2. Reported correctness is assessed separately from workflow status. Final outputs are evaluated under the corresponding benchmark protocol, including blinded mathematician judgments, LLM judging, or Lean kernel verification. Marking an issue resolved therefore does not itself count as a successful proof. The revised text now makes this distinction explicit.

AuthorsOpenReview ↗

Response to Reviewer QjBG (part 2/2)

Q1. Could a simpler operation combination achieve comparable performance?

Thanks for your insightful question. Section 4.2 evaluates simpler configurations and now also includes our prompt- and backbone-sensitivity experiments, with detailed settings and prompt-transformation instructions in the appendix. The component ablations remove individual critic analyses, literature grounding, or meetings and coordinated revision. The graded comparisons test simpler memory, context construction, and issue ordering. Local repair without meetings and coordinated revision scores 6.4 versus 8.5 for full RMA; severity-only ordering scores 7.23 ± 0.15 versus 8.53 ± 0.15. GIR with RMA's output-selection rule scores 5.50 ± 0.10. The additional experiments show that RMA maintains scores of 8.41–8.62 across the tested prompt variants and outperforms GIR by 2.32–3.08 points across three backbones. Together, these comparisons support the full design over the tested simpler alternatives.

Q2. Inconsistent counts.

The B2 counts under our invited panel's blinded evaluation are 7 passes for ProofCouncil, 3 for UCLA Moonshot, and 3 for ChatGPT 5.5 Pro, as reported in Table 2. RMA passes 8/10 problems.

Figure 2(a) is intended to plot these same panel judgments against realized expenditure. We have replotted Figure 2(a) to avoid confusions.

Q3. Practical evidence for preserving obligations and reusing prior work.

Our project page provides an interactive example on First Proof B1, Problem 6 (ε-light subsets), showing how proof obligations and prior attempts enter subsequent research operations: https://anonymous1776research.github.io/RMA-ICLR2027/#interactive-demo.

  1. An applied patch does not make the remaining mathematical obligation disappear. In the walkthrough, an initial repair bounds the leverage increase but leaves the barrier potential uncontrolled. The next critic round identifies this remaining gap in the good-vertex argument and records it as a new issue.

  2. The next repair uses the earlier attempt and its diagnosis. The solver receives the previous attempt, the current proof, and relevant concept records. The revised argument addresses both the leverage and barrier requirements. Meeting and revision operations also use accumulated attempts and literature notes to carry the barrier invariant through the induction.

  3. The example complements the controlled evaluation. The delivered proof establishes the constant 1/42 and was judged correct by all five experts. The intermediate records on the project page provide an illustrative reconstruction. The separate summary-memory comparison provides the empirical evidence: under the same harness and budget, full RMA achieves 25/30 blinded human passes versus 16/30. Together, the walkthrough explains how the mechanisms operate, while the controlled comparison measures their effect on final proof quality.


Overall, we have strengthened the comparison with concurrent systems through the Danus experiment and revised Related Work, and added prompt- and backbone-sensitivity results to Section 4.2. We also clarify the roles of solver and revision, explain the meaning of issue resolution, and provide a concrete project-page walkthrough of how proof obligations and prior attempts inform subsequent repairs. Together, these results and clarifications make RMA's methodological differences and practical effectiveness more explicit. If these results and clarifications address your concerns, we would greatly appreciate it if you would consider raising your score.

References:

[1] Liu, J., et al. (2026). Danus: Orchestrating Mathematical Reasoning Agents with Fact-Graph Memory. arXiv:2607.06447. https://arxiv.org/abs/2607.06447

[2] Schmitt, J., et al. (2026). ProofCouncil: An LLM Agent for Solving Open Mathematical Problems. arXiv:2607.09474. https://arxiv.org/abs/2607.09474

[3] Cao, Y., et al. (2026). MechMath Agent Team: LLM Driven Agents for Mathematical Research. arXiv:2607.04394. https://arxiv.org/abs/2607.04394

Reviewer_QjBGOpenReview ↗

Response to the Rebuttal

I greatly thank the authors for their detailed rebuttal. The authors added comparisons and discussions with additional baselines, and clarified the rationality and effectiveness of each module, which has addressed most of my concerns. Overall, I believe that persistent research state is not a novel contribution to AI4Math, but the context orchestration and some of the introduced components offer valuable inspiration. Taken together, I decided to maintain the current positive score.

We sincerely thank the reviewer for their thoughtful follow-up and for acknowledging that our rebuttal and additional experiments have addressed most of their concerns. We also greatly appreciate the reviewer's continued positive assessment and recognition of the value of our context orchestration and architectural components.

We would like to briefly reiterate the novelty of our contribution. We agree that persistent research state, as a general concept, has been explored in prior AI4Math systems. However, our contribution is not simply the introduction of persistent memory, but a systematic architecture that connects structured research state to long-horizon proof development through explicit context orchestration. In particular, RMA introduces:

  1. Proof-obligation-centered research state, explicitly linking unresolved issues, proof revisions, prior attempts, literature evidence, and evaluations to preserve the evolving structure of mathematical research.
  2. Research Context Orchestration, which dynamically prioritizes proof obligations, retrieves relevant research artifacts, and constructs bounded, operation-specific contexts to guide subsequent research actions.
  3. Coordinated local and global proof refinement, combining anchored issue-specific repairs, meeting-guided revisions, and repeated critique to preserve established progress while addressing remaining gaps.

Importantly, these contributions are supported by controlled experiments rather than architectural descriptions alone. Our additional comparison with Danus under matched experimental conditions, together with memory ablations and repeated experiments, provides further evidence that the way research state is organized, retrieved, and translated into actions is consequential for long-horizon reasoning performance.

We fully respect the reviewer's assessment of novelty and sincerely appreciate their constructive feedback and continued support. We hope this clarification further highlights the methodological contribution of RMA beyond the general concept of persistent research state.

Thank you again for your time and thoughtful engagement.

Reviewer_N7qJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a framework for constructing mathematical proofs for long-horizon tasks, claiming to address failure modes in which the agent is unable to maintain and expose the correct semantic state across many iterations. These failure modes leave certain proof-building claims unverified and potentially incorrect. Through the introduction of a state management layer and persistent research store, the proposed approach outperforms common benchmarks in the recent literature.

Critical strengths

The paper is well written, and the proposed approach is compared with the existing recent literature. The ablation tests in Section 4 are helpful in understanding the significance of the proposed state management layer.

Critical weaknesses

The empirical nature of the experimentation makes it difficult to assess the significance of the contribution, as it is not clear how variations of e.g., the prompt or the underlying LLM, would change the outcome of this work.

For the architecture investigation, why was the baseline of generic iterative revision chosen? The baseline seems rather simplistic in comparison with the full model. It is noted that other architectures were compared with in the Formal Conjectures component, but these were not budget-matched.

A large component of the paper is contingent on the retrieval mechanisms, but it is presently not explained exactly how this mechanism works.

Questions for authors

Below are some additional questions on the paper:

  • How dependent are the results on the prompts used?
  • Would the results change if applied to larger and more capable models such as fable?
  • Why are the models chosen for evaluation between B1 and B2 different? Why was there no measure on cost component for B1, and on formal proofs? Minor:
  • GIR table perhaps should be moved out of the appendix. Why was total cost chosen the way it was for this?
  • Graph 2(d) could be more explanatory.
3 comments
AuthorsOpenReview ↗

Response to Reviewer N7qJ (part 1/3)

We thank you for recognizing the quality of the presentation, the comparisons with recent systems, and the usefulness of the Section 4 ablations.

W1. Dependence on prompts and underlying models.

  1. About prompts. Prompts can affect both absolute performance and the size of RMA's improvement.

    (i) Section 4.2 provides evidence on how the choice of prompted operations affects performance. The “Component ablations” paragraph and Figure 2(c) examine removing individual critic analyses and removing meetings together with coordinated revision. These interventions remove the corresponding instructions and model calls. The “Comparisons with Generic Iterative Revision” paragraph also compares RMA with GIR, which uses generic critique-and-revise instructions instead of the specialized workflow. These results help assess the usefulness of the prompted operations at a high level, although they change the workflow as well as the instructions.

    (ii) To examine prompt wording more directly, we additionally conduct a prompt-sensitivity experiment on the same 30 held-out ResearchMath problems used in Section 4.2. We have included these results in Section 4.2 and provided the prompt-generation procedure and exact transformation instructions in the appendix. Motivated by studies of meaning-preserving prompt variations by Sclar et al. (2024) and Chatterjee et al. (2024), we compare three prompt sets: the original operation instructions of RMA in the paper, a paraphrased version, and a reformatted version. Each original system or task instruction is transformed separately, preserving its role and position in the workflow. The two alternative sets are generated offline with Claude Opus 4.8 using a fixed transformation prompt. For paraphrasing, we instruct it: “Rewrite this operation prompt using different wording while preserving every task requirement, constraint, output field, and template variable. Do not add hints or remove instructions.” For reformatting, we instruct it: “Keep the wording and requirement order unchanged, but organize the instructions using headings and bullet points. Preserve output schemas and template variables verbatim.”

    Each prompt set is evaluated with full RMA over three independent runs. We retain the settings of the existing held-out ablations: Claude Opus 4.8, the same reasoning settings and tools, a six-round limit, at most five issue repairs per round, and a 60,000-token context limit. Issue ranking, stopping rules, output selection, and RMA's memory and context-construction procedures remain unchanged across prompt variants. The evaluator operation described in Section 3.2 retains its original instructions, and anonymized final proofs are scored using the same three-model judge panel and rubric as the existing ablations. We report mean scores and standard deviations across runs and the observed score range across prompt sets. This experiment measures how RMA's proof quality changes under meaning-preserving variations in operation-prompt wording and formatting. The results are in the following table:

    Table A. Sensitivity of RMA to prompt variations

    Operation promptsRMA score ↑Difference from original
    Original8.53 ± 0.150.00
    Paraphrased8.41 ± 0.12−0.12
    Reformatted8.62 ± 0.16+0.09

    As Table A shows, RMA's performance remains stable across the tested prompt variants: the original, paraphrased, and reformatted instructions achieve 8.53 ± 0.15, 8.41 ± 0.12, and 8.62 ± 0.16, respectively. The mean scores span only 0.21 points on the 0–10 scale. These results suggest that RMA's measured performance does not rely on the particular wording or formatting of the original instructions.

  2. About underlying LLM backbones. For fair architectural comparisons, RMA and its controlled baselines use the same backbone, tools, and monetary allowance. Section 4.2 uses Claude Opus 4.8 for these comparisons, so differences in model capability do not explain the architectural gains. As documented in the “Models and inference” appendix, First Proof B1 was conducted earlier with Opus 4.6, used by both RMA and its standalone Claude baseline; B2 and the later experiments use Opus 4.8, except for the backbone-sensitivity study.

    We additionally test backbone dependence using Claude Opus 4.6, Claude Opus 4.8, and Fable 5.1 on the same 30 held-out ResearchMath problems used in Section 4.2, with three runs per configuration. We have included these results in Section 4.2 and documented the experimental controls in the appendix.

AuthorsOpenReview ↗

Response to Reviewer N7qJ (part 2/3)

Table B. Backbone sensitivity under matched experimental settings

BackboneRMA score ↑GIR + matched output selection score ↑Mean difference: RMA − GIR ↑
Claude Opus 4.67.21 ± 0.214.89 ± 0.14+2.32
Claude Opus 4.88.53 ± 0.155.50 ± 0.10+3.03
Fable 5.18.61 ± 0.195.53 ± 0.14+3.08

W2. Why use Generic Iterative Revision?

GIR tests whether a generic critique-and-revise loop can obtain the same benefit under the same monetary allowance. It has access to the same tools and can continue revising until its stopping condition or budget is reached. We use complementary controls to separate the contribution of structured memory from that of the broader workflow.

We choose GIR because iterative critique and revision provide a natural alternative for long-horizon proof development. GIR repeatedly examines and improves its current proof, with access to the same backbone, tools, and monetary allowance as RMA. It can therefore use its budget for successive repair attempts. RMA organizes this process through persistent research records, explicit issue prioritization, and operation-specific context construction. The comparison tests the value of this organization relative to a generic revision loop. We additionally equip GIR with RMA's output-selection rule for a more convincing comparison, allowing GIR to return its best-ranked proof rather than relying solely on its final revision.

The Formal Conjectures comparison uses a common monetary ceiling of $10 per target for all methods, alongside the same 300 targets and pinned Lean environments. The ceiling covers all attempts for each target, while each system retains its native inference procedure. We clarify this budget constraint in Section 4.1 and the implementation appendix.

W3. Retrieval mechanisms.

We have clarified the retrieval procedure in Section 3.2 (“Context compilation”) and expanded its implementation details in the appendix (“Additional Orchestration Mechanism Detail”). The revised explanation specifies how records are selected, ordered, and fitted into each operation's context.

  1. Selection follows explicit references between stored items. For example, a solver query identifies the issue being repaired. Retrieval returns that issue, the items it references, and items that reference it, such as literature notes and previous proof revisions addressing the gap. It follows direct links in both directions without expanding further. This is a deterministic lookup, without keyword matching, embeddings, or LLM-based selection.

  2. Ordering and context limits are explicit. Linked items follow a fixed priority order by record type, with earlier creation time breaking ties. The orchestrator also selects up to eight recent outputs from the same operation and removes duplicates. It assembles the instructions, current task, latest proof, linked items, and recent outputs, then removes whole items from the end when the context exceeds its budget. The appendix gives the exact priorities and recency rules.

  3. Retrieval occurs throughout the workflow. Each explicit Run invocation constructs an operation-specific context. With up to five selected issues, a complete round contains at most ten such constructions, or nine when the literature operation is disabled. These internal store reads are distinct from external literature search, which gathers sources and saves notes for subsequent operations.


Q1. How dependent are the results on the prompts used?

Answered in W1.

Q2. Would the results change with a larger model such as Fable?

Answered in W1 and Table B. With Fable 5.1, RMA scores 8.61 ± 0.19, compared with 5.53 ± 0.14 for GIR using RMA's output-selection rule under the same backbone and budget.

AuthorsOpenReview ↗

Response to Reviewer N7qJ (part 3/3)

Q3. Why do B1 and B2 use different models, and why are some costs not reported?

Thanks for your thoughtful question and constructive suggestions.

The model difference follows the timing of the benchmarks and experiments. B1 was run earlier with Opus 4.6 before B2 was released. Both RMA and its standalone Claude baseline use that backbone to ensure fair comparisons. B2 and the later experiments use Opus 4.8, except for the backbone-sensitivity study. We do not interpret differences between B1 and B2 as architectural improvements across model versions.

We have supplemented RMA's resource reporting in all benchmarks we used. The “Cost accounting” paragraph in the implementation appendix now reports total API costs of \1,826forFirstProofB1∗∗and∗∗1,826 for First Proof B1** and **\\2,738 for the 300 Formal Conjectures targets. These are realized expenditures, rather than budget ceilings. For B2, Section 4.1 and Figure 2(a) already report $1,850 for the headline run and $3,571 for the separate higher-budget run. For SOOHAK, RMA and the compared baselines were evaluated under the same $4,000 budget ceiling by the organizers.

Q4-5. GIR table and Figure 2(d).

  1. We have moved the GIR comparison table into the main text and added a per-problem cost column. Section 4.2 now explicitly references the table and explains the budget choice: \185perproblemmatchesRMA′sheadlineB2expenditureof185 per problem matches RMA's headline B2 expenditure of \\1,850 across ten problems. GIR receives the same backbone, tools, and monetary allowance, enabling a fair comparison of the two workflows.

  2. We have expanded both the Figure 2(d) caption and its explanation in Section 4.2. The “Graded design choices” paragraph now defines each memory, context-handling, and issue-ordering variant. The caption clarifies that these are three separate comparisons with a common full-RMA reference, and links directly to the main-text definitions.


Overall, the added experiments directly address sensitivity to prompts and backbones: RMA's mean scores span 0.21 points across the tested prompt variants, and it outperforms GIR by 2.32–3.08 points across three backbones. We also clarify the retrieval procedure and budget settings, expand resource reporting, move the GIR table into the main text, and define the Figure 2(d) variants explicitly. These revisions strengthen the empirical evidence and make the architecture and evaluation easier to assess. If these results and clarifications address your concerns, we would greatly appreciate it if you would consider raising your score.

References:

  1. Sclar, M., Choi, Y., Tsvetkov, Y., & Suhr, A. (2024). Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design or: How I Learned to Start Worrying About Prompt Formatting. ICLR 2024. Paper.

  2. Chatterjee, A., Renduchintala, H. S. V. N. S. K., Bhatia, S., & Chakraborty, T. (2024). POSIX: A Prompt Sensitivity Index for Large Language Models. Findings of EMNLP 2024, 14550–14565. Paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary and Manuscript Revisions

Dear AC and reviewers,

We sincerely thank all four expert reviewers for their thoughtful feedback, constructive suggestions, and positive scores. We especially appreciate their recognition of our work during a challenging ICLR review cycle, as the research community navigates significant changes and challenges. We have carefully addressed the reviewers' concerns through substantial new experiments, additional comparisons, and extensive revisions. We have revised and condensed the main paper to 10 pages, incorporating the key new results and clarifications directly into the manuscript. Here are the summaries of revisions:

1. New experiments and stronger baseline comparisons. We conducted additional experiments to strengthen the empirical evidence for RMA's architectural contributions:

  • Prompt sensitivity: RMA achieves stable performance across original, paraphrased, and reformatted prompts, with mean scores ranging from 8.41 to 8.62.
  • Backbone generalization: Across Claude Opus 4.6, Claude Opus 4.8, and Fable 5.1, RMA consistently outperforms budget-matched Generic Iterative Revision (GIR) by 2.32–3.08 points.
  • Comparison with Danus: Under matched backbone, tools, and budget, RMA achieves 8.53 versus 8.04 in continuous scores and 25/30 versus 20/30 passes in blinded mathematician evaluation.
  • Repeated experiments: We conducted three independent runs for six key configurations, reporting mean scores and standard deviations. The additional human evaluation further supports the contribution of persistent research state (25/30 versus 16/30 passes for summary memory).

2. Clarified methodology and architectural contributions. We substantially expanded the explanations of the Research Context Orchestrator, including deterministic record retrieval, context construction, prioritization, and budget handling. We also clarified the distinct responsibilities of solver and revision operations, the interpretation of issue resolution, and the differences between RMA and concurrent research-state architectures.

3. Improved evaluation transparency and reproducibility. We expanded API cost reporting, clarified budget-matching protocols and baseline settings, moved the GIR comparison into the main text, and improved the presentation of ablations and figures. We also added a concrete interactive walkthrough demonstrating how RMA preserves unresolved proof obligations and reuses prior attempts across research iterations.

4. Strengthened presentation and discussion. We revised the abstract to avoid overstating the causes of long-horizon reasoning failures, expanded Related Work and the discussion of broader applicability, and addressed the reviewers' questions regarding evaluation protocols, human inspection, and implementation details.


We are particularly grateful to the three reviewers who have now followed up, acknowledged that our rebuttal and revisions have addressed most of their concerns, and maintained their positive evaluations. We especially appreciate their recognition of the additional experimental evidence, improved methodological clarity, and strengthened evaluation transparency.

We welcome any further questions or concerns and remain fully committed to responding promptly and providing additional clarifications as needed.

Once again, we sincerely thank all four reviewers and the Area Chair for their time, expertise, and constructive engagement throughout the review process.

Best,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.