💻 computer science

MECA: A Mechanism-Centered Agent for Constructing Well-Specified and Valuable Mathematical Conjectures

The paper introduces MECA, a multi-agent framework that leverages mechanism-centered reasoning to transform broad research directions into well-specified, valuable, and challenging mathematical conjectures by jointly developing candidate statements and their underlying supporting mechanisms.

Original authors: Wentao Long, Yunfei Zhang, Chenyi Li, Zaiwen Wen

Published 2026-07-31

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Wentao Long, Yunfei Zhang, Chenyi Li, Zaiwen Wen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking for clues in a crime scene, you are looking for clues in the vast, dusty library of mathematics. For a long time, computers have been great at solving math problems that humans give them, like checking a specific homework assignment or proving a known theorem. But asking a computer to invent a new, interesting math problem is much harder. It's like asking a robot to write a mystery novel: if you just tell it to "write a story," it might produce something that makes no sense, is too vague, or has already been written a thousand times. The real challenge is getting the computer to come up with a question that is precise enough to be solved, but hard enough to be worth solving, and to explain why it thinks the answer might be "yes" or "no."

This is where the concept of a "mechanism" comes in. Think of a mechanism not as a physical gear, but as a specific, reusable trick or logic path that mathematicians use to connect a set of starting facts to a conclusion. It's like a specific recipe for baking a cake: if you have flour, eggs, and sugar (the assumptions), and you follow the mixing and baking steps (the mechanism), you get a cake (the conclusion). The problem is that if you just guess a new cake recipe without checking if the ingredients actually work together, you might end up with a brick. To build a good math conjecture, you need to find a new recipe and simultaneously verify that the ingredients actually support the steps.

Enter MECA (Mechanism-Centered Conjecture Agent), a new AI system designed to act like a team of super-smart, slightly obsessive math detectives. Instead of just guessing a math problem and hoping for the best, MECA works by building the problem and its supporting logic together, step-by-step. It uses a team of digital agents: some act as "Explorers" who try out different logic tricks and see if they fit, while others act as "Critics" who ruthlessly check if the logic holds up or if the problem is too easy or already solved.

The paper shows that MECA is surprisingly good at this. When tested on a challenge where it had to reconstruct a hidden math conclusion from old, incomplete notes (without seeing the answer key), MECA did a much better job than a standard AI that just guesses and edits. It managed to recover the precise details of the math problem, including the tricky conditions and the exact strength of the claim, far more accurately than the competition.

Furthermore, the team used MECA to generate 100 brand-new, semi-open math problems based on existing literature. They then handed these problems to a separate, powerful automated math solver (called QED) to see if it could solve them. The results were fascinating: MECA's problems were well-constructed and precise. About 35% were solved, 11% were proven wrong (which is a success because it means the problem was clear enough to be disproven), and the remaining 54% were too hard for the current solver to crack. This suggests that MECA isn't just making up gibberish; it's creating genuine, challenging mathematical puzzles that sit right on the edge of what current computers can handle. It turns broad, vague research ideas into sharp, well-specified questions that have a clear "unresolved core," making them perfect for the next generation of math discovery.

Technical Summary: MECA – A Mechanism-Centered Agent for Constructing Well-Specified and Valuable Mathematical Conjectures

1. Problem Statement

The automatic construction of well-specified and valuable mathematical conjectures remains a central challenge in AI-assisted mathematical discovery. Existing approaches often generate problems that are too broad, underspecified, or disconnected from plausible proof or refutation strategies. Current systems typically operate within predefined search spaces or begin from already selected research targets, lacking a general method to construct bounded conjectures from broad mathematical sources while jointly refining them with supporting mechanisms.

The core difficulty lies in the joint refinement of a candidate statement and its supporting reasoning. Fixing a statement too early may preserve an ill-posed target, while adapting it too freely to available arguments may render it trivial or already resolved. The goal is to produce conjectures that are precise, supported by plausible proof or refutation routes, valuable under either outcome, and retain a genuine unresolved core.

2. Methodology: The MECA Framework

The authors introduce MECA (MEchanism-centered Conjecture Agent), a multi-agent framework that treats candidate statements and their supporting mathematical mechanisms as coupled search objects.

2.1 Core Concept: Mathematical Mechanisms

A mechanism is defined as a reusable mode of reasoning (e.g., inequalities, invariants, decompositions, reductions) that links a set of assumptions to a corresponding conclusion. Unlike a vague proof technique, a mechanism in MECA is a concrete conditional argument characterized by:

$Req(m)$: The assumptions or previously established facts required.
$Prov(m)$: The consequences established by the argument.
$Gap(m)$: The non-routine conditions or downstream obligations that remain unresolved.

Mechanisms are organized into a Mechanism Graph ( $G_t$ ), a directed graph where edges represent dependencies (i.e., the conclusion of one mechanism supplies a condition required by another). This graph records how arguments compose, ensuring mutual compatibility and non-circularity.

2.2 The Three-Stage Pipeline

MECA operates through three distinct stages:

Source-Grounded Seed Construction:
- Input: Structured mathematical literature or manually curated broad open problems.
- Process: A Seed Generator converts materials into a "seed" recording the closest known result, a proposed advance, a preliminary target, and anticipated bottlenecks.
- Validation: A Seed Validator reviews for research value (rejecting restatements or trivial variations) and mathematical validity (checking definitions, quantifiers, and boundary cases). Only validated seeds proceed.
Mechanism–Problem Co-Exploration:
- Agents:
  - Mechanism–Problem Explorer: Proposes proof/refutation mechanisms, revises candidate statements based on evidence, and splits candidates if necessary.
  - Mechanism Critic: Independently audits mechanisms for correctness, non-circularity, and applicability.
  - Candidate Critic: Reviews candidate versions for well-posedness and alignment with the "frozen target anchor" (the fixed research boundary).
  - Coverage Judge: Evaluates whether accepted mechanisms compose to form a coherent route toward the candidate, distinguishing between Insufficient Support, Substantive Partial (meaningful support with explicit unresolved cores), and Full Coverage.
- Process: The candidate statement and mechanism graph are refined jointly. If mechanisms reveal missing assumptions or an overly strong conclusion, the candidate is revised.
Prover-Facing Problem Construction:
- Final Challenger: An isolated agent that audits the candidate without access to the mechanism graph or exploration history. It checks for statement closure, counterexample resistance, and value.
- Projection: Accepted candidates are deterministically projected into self-contained problem statements ( $q = \Theta(A, p, G')$ ). Unresolved cores are converted into concrete proof obligations.
- Blind Probes: The final statement is tested by independent provers (without access to MECA's internal reasoning) to assess difficulty and solvability.

3. Key Contributions

The paper makes three primary contributions:

Formulation: It formulates conjecture construction as the joint refinement of a candidate statement and its supporting mechanisms, guided by audited partial support and explicit unresolved obstacles.
Framework: It introduces MECA, a mechanism-centered multi-agent framework capable of constructing precise conjectures from mathematical literature and broad open problems.
Dataset: It contributes 100 prover-facing open problems with explicit assumptions and mathematical motivation, designed to expose the limitations of current automated proof and refutation systems.

4. Experimental Evaluation

MECA was evaluated in two complementary settings:

4.1 Target-Conditioned Conclusion Recovery

Setup: The system was tasked with reconstructing conclusions from 20 post-2023 target papers using only pre-2024 "article-blind" source materials (seven papers per target).
Baseline: Compared against a "Draft–Audit–Revise" baseline.
Results: MECA significantly outperformed the baseline, improving the average target-alignment score from 48.0 to 69.0. The largest gains were in recovering the central conclusion (+6.0) and quantitative accuracy (+5.5), indicating that mechanism-centered construction improves the recovery of precise theorem-level claims and their validity regimes.

4.2 Conjecture Generation and Verification

Setup: 100 semi-open problems were constructed from literature seeds and existing open problems. These were submitted to QED, an open-source multi-agent theorem prover, with a 5-hour time limit.
Results:
- 35% were proved.
- 11% were refuted (verified counterexamples).
- 54% remained unresolved.
Analysis of Unresolved Cases: Among the 54 unresolved cases, 81.5% were attributed to technical proof difficulty (the statement was well-posed and supported, but the proof exceeded current solver capabilities). Only 10% of all generated conjectures exhibited formulation or proof-readiness defects (e.g., missing assumptions, ambiguity). This suggests MECA produces conjectures that are usable as standalone tasks while retaining substantive mathematical difficulty.

5. Significance and Claims

The paper claims that MECA successfully transforms broad research directions into precise conjectures with substantive mathematical support while retaining a clearly identified unresolved core. By organizing refinement around mathematical mechanisms, the framework ensures that:

Candidates are not merely speculative but are supported by audited partial progress.
The "unresolved core" is explicitly isolated, distinguishing between a lack of support and a genuine open problem.
The resulting conjectures are challenging for current automated provers, as evidenced by the high rate of "technical difficulty" outcomes rather than "formulation defects."

The authors emphasize that MECA does not claim to solve these problems or guarantee historical novelty in an absolute sense, but rather demonstrates a reliable method for constructing well-specified, research-worthy conjectures from bounded mathematical evidence. The framework addresses the gap between generating isolated statements and developing them in tandem with the structural reasoning required to validate them.