Skip to content

README.md

Latest commit

 

History

History
141 lines (98 loc) · 5.76 KB

README.md

File metadata and controls

141 lines (98 loc) · 5.76 KB
Raw

Math Modeling Multi-Agent Paper Generation System

Stars Version License: MIT

Python LaTeX Claude Code CUMCM / MCM / HiMCM

中文 ← current · English · 日本語 · 한국어

v2.2 — Multi-agent adversarial collaboration + 9 independent QA gates + anti-pattern hard blocks + human-in-the-loop checkpoints + cross-stage state management. 21 reference manuals, 6 algorithm cookbooks, bilingual phrase bank, anti-pattern knowledge base, environment doctor script.

Drop in a contest problem PDF, and the system automatically runs through: problem analysis → method selection → web data hunting → Python solving → visualization → LaTeX paper compilation → 3-role parallel adversarial review → final PDF. Supports CUMCM (China), MCM/ICM (USA), and HiMCM.

Quick Start

Prerequisites

  • Node.js ≥ v18 — nodejs.org
  • Claude Code: npm install -g @anthropic-ai/claude-code

Install

git clone https://github.com/Linference/math_model.git ~/.claude/skills/math-modeling

Or download the ZIP from Releases.

Run

claude
/math-modeling
[paste contest problem or drag in PDF]

7-Stage Pipeline

Stage Task Output
0 Project scaffolding Standard directory skeleton
1 Deep problem analysis Structured analysis report (2000+ words)
2 Method selection + ML/DL decision Modeling plan with chart checklist
3 Web data hunting CSV files + SOURCES.md
4 Python implementation Runnable scripts + numerical results
5 Visualization (16 chart types) High-quality figures (300 DPI)
6 LaTeX writing + compilation Paper draft
7 3-role adversarial review Final PDF (score ≥ 7.5/10)

8 Sub-agents

Agent Role
mm-problem-analyst Problem analysis: decompose questions, find constraints & traps
mm-modeler Method selection: ML/DL decisions, chart planning
mm-data-hunter Data: search Wikipedia/GitHub/Kaggle/sklearn
mm-coder Implementation: Python solving + visualization
mm-writer Writing: fill LaTeX template, address review feedback
mm-reviewer Review: 5-dimension scoring, find weaknesses
mm-verifier Verification: cross-check numbers, units, boundary conditions
mm-reasoner Reasoning: audit formula derivations, fill proof gaps

Adversarial Review

Three reviewer roles evaluate the same paper in parallel — Reviewer (modeling quality), Verifier (numerical accuracy), Reasoner (mathematical rigor). The Writer addresses each issue, then all three re-score. Loop continues until average score ≥ 7.5 or max 4 rounds.

Quality Gates (v2.2)

9 independent Subagent QA gates + anti-pattern hard blocks + human-in-the-loop checkpoints ensuring output quality at every stage.

Project Structure

math_model/
├── README.md
├── samples/                      # Full example projects
│   ├── 2024_CUMCM_A/             # CUMCM Problem A
│   └── 2025_HiMCM_Problem_B/     # HiMCM Problem B
└── skill/                        # Skill installation package
    ├── SKILL.md                  # Main skill definition
    ├── references/               # 15 manuals + 6 cookbooks
    ├── scripts/                  # Helper scripts
    ├── templates/                # LaTeX templates (CN/EN)
    └── workflows/                # Review workflows

Sample Projects

2024 CUMCM A — Dragon Bench Motion

Model the motion trajectory of 223 bench sections along an Archimedean spiral, detect collisions, and optimize turnaround paths. Techniques: chord-length constrained binary search, SAT collision detection, constrained nonlinear optimization.

2025 HiMCM B — Super Bowl Sustainable Host City Selection

Build an environmental-only site selection model for NFL, evaluate 19 historical + 3 candidate cities, extend to Olympics. Techniques: AHP + TOPSIS two-tier decision making, Scope 1/2/3 carbon emission analysis, sensitivity analysis.

Changelog

v2.2.0 — Deep Fix (current)

  • Bug fix: scoring aggregation upgraded from simple average to role×dimension weighted matrix + hard caps + outlier arbitration
  • Anti-pattern hard blocks (A1/A2/A3): code/writing/full-dimension anti-pattern scans, High hits must be fixed
  • Data quality gate (D1): missing values + outliers + multi-source alignment checks at stage 3
  • Human-in-the-loop checkpoints: pause after stages 1/4/6 for user confirmation
  • Gates expanded from 5 to 9

v2.1.0 — Knowledge Embedding + Literature Review

  • 8 agent definitions fully upgraded (thin→thick), embedding domain knowledge
  • New stage 1.5: literature review (required for national first prize)
  • Enhanced innovation validation: ablation studies + 3 baseline comparisons + multi-scenario validation

v2.0.0 — Architecture Upgrade

  • Independent Subagent QA protocol (5 gates)
  • Cross-stage state management
  • Reference manuals major upgrade (6 manuals, 3-13x length growth)

v1.2 — Initial Release

  • 7-stage mandatory pipeline + 8 sub-agents + 3-role adversarial review

License

MIT