Technical Summary: RMA – An Agentic System for Research-Level Mathematical Problems

Problem Definition

Current benchmarks for mathematical reasoning in artificial intelligence, such as MATH or International Mathematical Olympiad (IMO) problems, are limited to well-scoped competition-style tasks. These problems typically admit concise, self-contained solutions using known techniques and do not require engagement with external literature or the development of novel proof strategies. In contrast, research-level mathematical problems present a qualitatively different challenge. They require systems to interpret specialized definitions, formalize assumptions, identify relevant prior results from vast literature, explore effective proof strategies, construct intermediate lemmas, and execute long chains of rigorous reasoning. Existing formal theorem proving systems are often ill-suited for this domain because they assume fully specified problems and fixed formal contexts, whereas research mathematics demands engagement with unstructured literature and constructive techniques. Furthermore, existing agentic research tools are not designed for the structured reasoning, gap identification, and rigorous presentation required for mathematical proofs.

Methodology: The RMA Framework

The authors propose Research Math Agents (RMA), an agentic framework designed specifically for automated reasoning on research-level mathematical problems. RMA decomposes the proof-solving process into specialized modules and coordinates them through a multi-agent, multi-round workflow.

1. Modular Design

RMA is built upon six specialized modules implemented as structured routines executed by CLI-based coding agents:

• Problem Analysis Module: Transforms raw problem statements into structured intermediate descriptions by formalizing variables, decomposing problems into subgoals, and extracting constraints. This reduces ambiguity and guides subsequent reasoning.
• Literature Search Module: Retrieves relevant prior work by generating candidate queries based on the structured problem representation rather than simple keyword matching. It handles variability in mathematical notation across platforms (e.g., arXiv, MathSciNet) and employs a "candidate generation" step to prevent information leakage from existing solutions during retrieval.
• Fair Comparison Module: Ensures evaluation integrity by filtering out sources containing known solutions, isolating interaction history between runs, sandboxing tool use, and utilizing foundation models with training cutoffs predating the benchmark release to prevent data contamination.
• Literature Understanding Module: Processes retrieved literature by extracting candidate lemmas and techniques, filtering them for alignment with the current problem's assumptions and targets, and organizing them into summaries grouped by role (e.g., supporting subgoals).
• Knowledge Bank Module: A "cheat-sheet" style repository of concise, reusable entries (e.g., standard inequalities, combinatorial identities) curated from canonical textbooks. These entries include explicit assumptions and applicability conditions, allowing the system to instantiate and verify standard arguments without re-deriving them.
• Proof Commandment Module: Enforces a set of rules to mitigate common LLM failure modes, including:
  • Grounding: Claims must be supported by established results or derivations.
  • Faithfulness: Adherence to the original problem statement.
  • Gap-free reasoning: Explicit justification for every inference.
  • Constructiveness: Requirement for explicit constructions where necessary.
  • Format correctness: Clean, compilable LaTeX output.

2. Multi-Agent and Multi-Round Instantiation

RMA operationalizes these modules through a system of role-specialized agents interacting via a Shared Structured Memory stored on disk. The system consists of:

• Initializer Agent (Single): Produces the initial proof draft, populates the memory with problem analysis, literature context, and reusable knowledge.
• Proposer Agents ($K_p$): Refine the evolving proof by identifying gaps, invoking relevant modules, and developing new arguments. They have write access only to the proof state.
• Verifier Agents ($K_v$): Evaluate the current proof against the Proof Commandment Module, checking logical consistency and identifying missing steps. They have write access only to the feedback state.

The workflow proceeds in $N$ rounds. In each round, proposers generate candidate proofs conditioned on the current memory state, and verifiers provide structured feedback. The shared memory contains the problem state, literature context, knowledge entries, proof state, and accumulated feedback. Read/write permissions are strictly controlled to prevent conflicts, with all updates appended with agent and round identifiers.

Key Contributions

1. RMA Framework: The proposal of an agentic framework that decomposes research-level proof solving into domain-specific modules (problem analysis, literature search, knowledge retrieval, and verification) while actively preventing data contamination.
2. Multi-Agent, Multi-Round Architecture: The introduction of a system that decomposes roles into proposers and verifiers, enabling iterative reasoning, automatic verification, and self-improvement through structured interactions, which improves robustness over single-agent approaches.
3. Empirical Validation: Rigorous expert verification demonstrating that RMA outperforms strong baselines (including GPT-5.2R and Aletheia) in solving research-level problems. The system achieves higher correctness rates and produces proofs that are more interpretable and aligned with human mathematical reasoning.
4. Ablation Insights: Systematic studies revealing that performance gains arise from the interaction of structured modules, iterative refinement, and verifier-based feedback, rather than any single component.

Experimental Results

The system was evaluated on the First Proof benchmark, consisting of ten research-level problems contributed by expert mathematicians across diverse domains (e.g., stochastic analysis, spectral graph theory).

• Correctness: RMA successfully solved 8 out of 10 problems according to expert judgment. This significantly outperforms strong baselines: GPT-5.2R (3/10), Aletheia (5/10), and Agentic Researcher (3/10). Retrieval-augmented research assistants (GPT-DR, Gemini-DR) failed to produce any fully correct solutions.
• Quality: In fine-grained evaluations (Answer Accuracy, Logical Correctness, Completeness, Clarity), RMA achieved the highest scores across all dimensions.
• Pairwise Comparison: In blind pairwise A-B comparisons, RMA ranked first with a win rate of 0.75 against experts and 0.78 against LLM evaluators, outperforming all baselines.
• Ablation Studies:
  • Removing the Problem Analysis or Knowledge Bank modules caused severe performance degradation.
  • Multi-agent reasoning significantly outperformed single-agent generation and compute-matched "best-of-N" sampling.
  • Performance peaked at 5 reasoning rounds ($N=5$); excessive iterations led to degradation.
  • Scaling the number of verifiers ($K_v$) and proposers ($K_p$) improved performance up to a point (3 agents each), after which diminishing returns or slight degradation were observed.
  • Structured memory (preserving history across rounds) was critical, outperforming stateless or "last-round only" memory designs.

Significance and Claims

The paper positions RMA as a use-inspired advancement in AI for scientific discovery. The authors claim that the framework demonstrates that agentic designs are a promising direction for advancing AI toward open-ended mathematical reasoning.

Key takeaways emphasized by the authors include:

• Beyond Competition Math: The shift from solving competition problems to assessing AI's ability to contribute to mathematical discovery requires a different architectural approach, specifically one that integrates literature grounding and iterative refinement.
• Interaction over Components: The performance gains are attributed to the interaction of structured reasoning modules, iterative refinement, and verifier-based feedback, rather than the strength of any single component.
• Human-in-the-Loop: The system is designed as an assistant to help researchers explore strategies and organize literature, not as a replacement for expert judgment. The authors acknowledge that expert verification remains necessary to validate assumptions and citations, as LLMs can still produce convincing but incorrect arguments.
• Reproducibility: The modular design and shared memory structure make the reasoning process inspectable, allowing users to audit the origin of claims and the flow of the proof construction.

The authors conclude that while RMA represents a significant step forward, the field requires continued expansion of benchmarks, cost-accuracy trade-off analysis, and transparent evaluation protocols to fully realize the potential of AI in mathematical research.