yushui2022/MathModel-Skill

What it solves

MathModel Skill provides a structured framework for AI programming agents to handle the entire lifecycle of mathematical modeling competitions. Instead of relying on a single prompt to generate a paper, it transforms the process into a verifiable workflow that ensures actual calculations are performed, evidence is gathered, and a formal paper is produced in a professional format (Word/PDF).

How it works

The project implements a multi-stage pipeline (S0-S8) where different "Skills" handle specific tasks. The process moves from input pre-checking and problem analysis to model selection, coding, execution, and finally formal writing and rendering checks.

Key mechanisms include:

• Evidence-Based Workflow: Each step produces actual files and evidence (recorded via SHA-256) that the next step depends on, preventing the agent from hallucinating results.
• Tiered Capability Levels: It offers three versions (Lite, Standard, Pro) tailored to different LLM capabilities, ranging from basic reports to high-intensity research with multi-route comparisons and robustness experiments.
• Formal Authoring: A dedicated writer component creates the paper section-by-section, audits the content, and merges them into a final document with native Word formulas and PDF rendering via LibreOffice.

Who it’s for

Students or researchers participating in mathematical modeling competitions who use AI agents (such as Codex, Claude Code, or Trae) to assist in problem analysis, data processing, and paper drafting.

Highlights

• Verifiable Execution: Forces the agent to run real code and retain evidence rather than just simulating the process.
• Multi-Platform Support: Ready-made installation packages for Codex, Claude Code, and Trae.
• Format Rigor: Includes automated checks for paper length, native Word OMML formulas, and PDF rendering verification.
• Tiered Workflows: Flexible versions (Lite, Standard, Pro) to match the reasoning power and cost of the chosen LLM.