paper2code-enhanced

name	paper2code-enhanced
description	Paper2Code-Enhanced — multi-agent paper-to-code pipeline (planning → analyzing → coding) using PaperCoder. Generates working implementations from ML papers. ~$0.50–0.70 per run, ~1 hour per run.
type	skill-manual
created	"2026-05-24T00:00:00.000Z"
updated	"2026-05-24T00:00:00.000Z"
tags	["paper-to-code","ml","reproduction","papercoder","autonomous-agents"]
triggers	["paper to code","generate from paper","reproduce paper","paper2code","TreeOfThoughts","Yao et al"]

Paper2Code-Enhanced Skill Manual

Paper2Code-Enhanced transforms ML papers into code repositories via PaperCoder — a three-stage multi-agent pipeline (planning, analysis, coding).

Repo: /home/ty/Repositories/ai_workspace/Paper2Code-Enhanced/ Entry: codes/pipeline.py (Python API)

---

When to Use

• Paper has an official implementation you want to reproduce or study
• You want a working scaffold from a paper without reading it line-by-line
• Paired with OrCAID commit0 for validation+fixup
• Cost: ~$0.50–0.70 per full run (MiniMax M2.7)
• Runtime: ~1 hour per full run

Do NOT use for:

• Simple papers with obvious implementations (overkill)
• Real-time needs (1 hour per run)
• Papers without clear algorithmic specs

---

Python API (Recommended)

import sys
sys.path.insert(0, '/home/ty/Repositories/ai_workspace/Paper2Code-Enhanced/codes')

from pipeline import run_pipeline, PipelineConfig

config = PipelineConfig(
    paper_name="TreeOfThoughts",
    pdf_json_path="/path/to/arxiv_2605.15612_cleaned.json",
    output_dir="/home/ty/Repositories/ai_workspace/paper2code-projects/TreeOfThoughts/",
    output_repo_dir="/home/ty/Repositories/ai_workspace/paper2code-projects/TreeOfThoughts_repo/",
)
result = run_pipeline(config)
print(result.status, result.output_repo_dir)

Requires PYTHONPATH — always prepend the codes directory:

import sys
sys.path.insert(0, '/home/ty/Repositories/ai_workspace/Paper2Code-Enhanced/codes')

---

Three Stages

Stage	Script	What it does
Planning	codes/1_planning.py	Decomposes paper into modules, generates planning_artifacts/
Analyzing	codes/2_analyzing.py	Extracts algorithm details, produces analyzing_artifacts/
Coding	codes/3_coding.py	Generates the full repository in <PaperName>_repo/

Each stage writes results to SQLite DB (codes/db.py) and artifacts to output_dir/.

---

Input Preparation

From arXiv JSON (cleaned format)

cd /home/ty/Repositories/ai_workspace/Paper2Code-Enhanced

uv run python codes/0_pdf_process.py \
    --input_json_path examples/Transformer.json \
    --output_json_path examples/Transformer_cleaned.json \
    --mode auto

From PDF directly

uv run python codes/0_pdf_process.py \
    --input_json_path path/to/paper.pdf \
    --output_json_path output_cleaned.json \
    --mode auto

Modes: auto (default, VLM if key available, else local), vlm (Vision-Language Model), local (pypdf, offline), olmocr (olmOCR pipeline)

From LaTeX source

cd scripts
uv run bash run_latex.sh

---

Output Structure

outputs/
├── <PaperName>/           # Artifacts (planning, analyzing, coding artifacts)
│   ├── planning_artifacts/
│   ├── analyzing_artifacts/
│   ├── coding_artifacts/
│   └── accumulated_cost.json
└── <PaperName>_repo/      # Generated code repository (THE OUTPUT YOU WANT)
    ├── config.yaml
    ├── main.py
    ├── utils.py
    ├── tasks/
    └── tot/               # (for TreeOfThoughts paper)

---

Configuration (.env)

Set in /home/ty/Repositories/ai_workspace/Paper2Code-Enhanced/.env:

LLM_API_KEY=your_key_here
LLM_BASE_URL=https://api.minimax.io/v1
LLM_MODEL=MiniMax-M2.7

Base URL must be https://api.minimax.io/v1 (OpenAI-compatible endpoint). The /anthropic/v1/messages endpoint only serves Claude models and returns 404 for MiniMax.

MiniMax M2.7 recommended — supports thinking/reasoning blocks natively via Anthropic SDK.

---

CLI (Broken — Use Python API)

The CLI scripts (scripts/run.sh, codes/1_planning.py, etc.) do from db import ... but don't set PYTHONPATH, causing import errors.

Always use the Python API for programmatic runs.

If calling stage scripts directly, prepend PYTHONPATH:

PYTHONPATH=/home/ty/Repositories/ai_workspace/Paper2Code-Enhanced/codes uv run python codes/1_planning.py ...

---

Common Issues

db module import error

CLI scripts do from db import ... but don't set PYTHONPATH. Use Python API instead.

run.sh empty --gpt_version

If LLM_MODEL env var is empty, GPT_VERSION="" gets passed as --gpt_version "" which scripts interpret as a positional argument. Ensure .env has LLM_MODEL=MiniMax-M2.7 set.

Missing _cleaned.json suffix

Pipeline expects <paper_name>_cleaned.json (with the _cleaned suffix). Running 0_pdf_process.py produces this.

---

OrCAID Integration

Paper2Code-Enhanced generates a repository; OrCAID commit0 validates and fixes it:

1. Paper2Code generates → <PaperName>_repo/ with implementation
2. OrCAID commit0 validates → runs pytest, parses JSON report, applies patches

cd /home/ty/Repositories/ai_workspace/OrCAID && \
uv run python -m orcaid.cli \
    --task=commit0 \
    --model=minimax/MiniMax-M2.7 \
    --multi_agent=false \
    --max_iterations=5 \
    --patch_target /home/ty/Repositories/ai_workspace/paper2code-projects/TreeOfThoughts_repo/

---

Key Files

File	Purpose
codes/pipeline.py	Python API entry (run_pipeline, PipelineConfig, PipelineResult)
codes/1_planning.py	Planning stage
codes/2_analyzing.py	Analyzing stage
codes/3_coding.py	Coding stage
codes/0_pdf_process.py	PDF/JSON ingestion + cleaning
codes/eval.py	Evaluation script
codes/db.py	SQLite persistence (sqlmodel)
examples/Transformer_cleaned.json	Example input

---

Constraints

• Always use Python API for programmatic runs (CLI broken due to db module path)
• Prepend PYTHONPATH when calling stage scripts directly
• LLM_MODEL must be set in .env before running
• ~1 hour per full run — don't use for trivial papers
• Cost ~$0.50–0.70 per run with MiniMax M2.7