MathModelAgent: an end-to-end automation agent for mathematical modeling and academic paper generation

What it solves

MathModelAgent automates the complex process of mathematical modeling for competitions. It aims to reduce the time required to complete a high-quality modeling paper from several days to approximately one hour by handling everything from problem analysis to final document formatting.

How it works

The project uses a multi-agent system (including specialized agents for modeling, coding, and writing) that can be powered by various LLMs via LiteLLM. It follows an end-to-end workflow:

1. Analysis & Modeling: Analyzes the problem and selects appropriate mathematical models.
2. Coding: Uses a Code Interpreter (local via Jupyter or cloud-based via E2B/Daytona) to implement models and generate figures.
3. Writing: Generates a formatted paper using Typst templates (supporting major competitions like MCM/ICM).
4. Verification: Performs a 9-step acceptance check, including leakage detection and numerical consistency checks.

Optional features include Web Search (Tavily API) for real-world data, a RAG knowledge base (ChromaDB) for modeling methods, and Human-in-the-Loop (HIL) checkpoints for user approval.

Who it’s for

Students and researchers participating in mathematical modeling competitions who want to automate the drafting, coding, and formatting of their submissions.

Highlights

• End-to-End Automation: Handles the entire pipeline from problem analysis to PDF generation.
• Typst Integration: Includes 17 professional paper templates for various international and national competitions.
• Flexible Execution: Supports both a standalone WebUI/CLI and integration as a "SKILL" for harnesses like Claude Code or Codex.
• Robust Tooling: Features a built-in code interpreter and a specialized knowledge base for model selection (e.g., AHP, TOPSIS, ARIMA).
• Error Handling: Implements a four-layer fault tolerance system including retries and fallback hand-offs.