SolveX

Multi-Agent mathematical modeling system based on OpenManus.

Overview

SolveX automates the full mathematical modeling pipeline for competition-style problems:

Problem → [Modeling ↔ Programming Loop] → Visualization → LaTeX Paper → Output

Agents:

• ModelingAgent — Analyzes problems, searches literature (ArXiv, web), designs innovative models
• ProgrammingAgent — Implements models in Python, verifies results
• VisualizationAgent — Creates publication-quality figures with matplotlib
• WritingAgent — Composes complete LaTeX papers

Key features:

• Iterative modeling-programming loop with automatic verification
• Literature search via ArXiv + web MCP tools
• Memory compaction for long sessions
• Data file support (CSV, XLSX, ZIP)
• FastAPI web UI with SSE streaming + ZIP download
• 20-minute end-to-end pipeline for competition problems

Quick Start

# Setup
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Configure LLM (edit config/config.toml)
cp config/config.example.toml config/config.toml

# CLI mode
python run_flow.py --problem tests/competition/2025_C/problem.txt --data tests/competition/2025_C/data/

# Web mode
python api.py  # open http://localhost:8000

Architecture

app/
├── agent/           # Agent implementations (modeling, programming, visualization, writing)
├── flow/            # Workflow orchestration (SolveXFlow)
├── prompt/          # System prompts for each agent
├── service/         # Memory compaction service
├── tool/            # Tools (python_execute, str_replace_editor, MCP, etc.)
├── llm.py           # LLM abstraction (supports DeepSeek, OpenAI, Bedrock, Azure, Ollama)
└── schema.py        # Data models
config/
├── config.toml      # LLM configuration
└── mcp.json         # MCP tool servers
static/              # Web frontend (HTML/CSS/JS)
api.py               # FastAPI server
run_flow.py          # CLI runner

Workspace Output

Each run produces:

workspace/
├── 01_modeling/     # plan.md, final_plan.md
├── 02_programming/  # solution.py, predictions, results
├── 03_figures/      # PNG figures, figures_catalog.md
└── 04_paper/        # main.tex (complete LaTeX paper)

License

MIT License — Based on OpenManus.