🧩 ProgramGym-1648

A dataset of 1,648 tasks for long-horizon program reproduction agents.

📄 Paper · 🌐 Project Page · 💻 GitHub

---

📖 Overview

ProgramGym-1648 is a collection of 1,648 Harbor tasks designed to train an AI agent's ability to reproduce software programs from their observed behavior.

In each task, the agent is given a reference executable and accompanying documentation. The objective is to independently reconstruct the program's functionality by developing a new codebase that matches the reference program's behavior.

The agent must deliver:

• 📝 Source Code — An original implementation of the target program.
• 🔧 Build Script — A compile.sh script that compiles the implementation.

🚫 Important: Copying, wrapping, or decompiling the reference executable is strictly prohibited. Agents must reproduce the program's functionality through their own implementations.

📂 Task Structure

Each task follows a standardized Harbor-compatible directory structure:

Path	Description
📄 task.toml	Harbor task metadata and runtime configuration
📋 instruction.md	Task requirements, instructions, and constraints
🖥️ environment/	Agent execution environment and reference executable
🧪 tests/	Verification logic and test fixtures
✅ solution/	Harbor solution entry point for validation

🚀 Quick Start

ProgramGym-1648 can be evaluated directly using Harbor.

1. Prerequisites

Before running the benchmark, make sure you have:

• 🐳 Docker installed and running.
• 🔑 A configured model provider with valid API credentials.
• 🛠️ Harbor ≥ 0.20.0, supporting task schema 1.3 and phase-scoped network policies.

Install Harbor using uv:

uv tool install 'harbor>=0.20.0'

2. Run the Benchmark

Launch the full benchmark with the following command:

harbor run \
  -d ProgramGym/programgym-1648 \
  -a claude-code \
  -m your-provider/your-model \
  -n 4

Configuration:

Argument	Description
-d	Dataset identifier
-a	Agent implementation (e.g., claude-code)
-m	Model provider and model identifier
-n	Number of concurrent trials (4 in this example)

Replace your-provider/your-model with your target model and configure the corresponding API credentials.

3. Results

📊 By default, Harbor writes execution logs and results to the jobs/ directory.

📚 Citation

If you find ProgramGym useful in your research, please consider citing our paper:

@misc{yang2026programgym,
  title         = {ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents},
  author        = {Chengyi Yang and Renxu Bao and Jue Hou and Zongpei Teng and Ziheng Wang and Yujie Lin and Chengsong Huang and Junnan Zhu and Delai Qiu and Yining Wang and Shengping Liu and Fei Long and Jinsong Su},
  year          = {2026},
  eprint        = {2610.13395},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2610.13395}
}

📜 License

This dataset is released under the MIT License.