ProgramBench: Can Language Models Rebuild Programs from Scratch?

Meta's Facebook Research released ProgramBench, a benchmark in which AI agents must architect and implement a complete codebase that reproduces a program's behavior given only its compiled binary and documentation. The benchmark is published as arXiv paper 2605.03546 by John Yang, Kilian Lieret, Jeffrey Ma, Parth Thakkar, Dmitrii Pedchenko, Sten Sootla, Emily McMilin, Pengcheng Yin, Rui Hou, Gabriel Synnaeve, Diyi Yang and Ofir Press, and ships as an installable Python package (uvx programbench, uv pip install programbench, or pip install programbench) with a mini-swe-agent baseline runnable via mini-extra programbench.

Can Language Models Rebuild Programs From Scratch? Given only a compiled binary and its documentation, AI agents must architect and implement a complete codebase that reproduces the original program's behavior. We recommend uv https://docs.astral.sh/uv/getting-started/installation/ for managing Python environments. Run without installing uvx programbench --help Or install into a project uv pip install programbench Or with pip pip install programbench For development: git clone https://github.com/facebookresearch/programbench.git cd programbench uv sync installs editable + dev dependencies Note For more details, please refer to the Usage Guide https://github.com/facebookresearch/ProgramBench/blob/main/docs/README.md . To run our baseline: uvx --from mini-swe-agent mini-extra programbench --help or pip install mini-swe-agent && mini-extra programbench --help See the mini-swe-agent baseline docs https://mini-swe-agent.com/latest/usage/programbench/ for details. If our work was useful for you, please cite it: @misc{yang2026programbenchlanguagemodelsrebuild, title={ProgramBench: Can Language Models Rebuild Programs From Scratch?}, author={John Yang and Kilian Lieret and Jeffrey Ma and Parth Thakkar and Dmitrii Pedchenko and Sten Sootla and Emily McMilin and Pengcheng Yin and Rui Hou and Gabriel Synnaeve and Diyi Yang and Ofir Press}, year={2026}, eprint={2605.03546}, archivePrefix={arXiv}, primaryClass={cs.SE}, url={https://arxiv.org/abs/2605.03546}, } ProgramBench is licensed under the terms of the license found in LICENSE https://github.com/facebookresearch/ProgramBench/blob/main/LICENSE .