@KLieret: You can evaluate on ProgramBench yourself: https://github.com/facebookresearch/ProgramBench/… We will open the leaderbo…
Summary
ProgramBench is a new benchmark that tests AI agents' ability to reconstruct a complete codebase from a compiled binary and its documentation. The leaderboard will open for submissions soon.
View Cached Full Text
Cached at: 05/14/26, 04:29 AM
You can evaluate on ProgramBench yourself: https://github.com/facebookresearch/ProgramBench/… We will open the leaderboard for submissions soon.
facebookresearch/ProgramBench
Source: https://github.com/facebookresearch/ProgramBench

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.
Links
Quickstart
We recommend uv 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
For more details, please refer to the Usage Guide.
Citation
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},
}
License
ProgramBench is licensed under the terms of the license found in LICENSE.
Similar Articles
ProgramBench (5 minute read)
ProgramBench is a new benchmark that evaluates AI agents' ability to reconstruct complete software projects from compiled binaries and documentation without access to source code or decompilation tools.
@Lyubh22: Coding benchmarks are saturating. AI4Research is the next frontier. Thrilled to see our MLS-Bench (https://mls-bench.co…
Announcing MLS-Bench, the first AI4Research benchmark to gain broad community adoption, testing AI agents on 140 executable tasks across 12 domains to propose modular ML improvements. The post includes leaderboard scores for models like Claude Opus 4.6 and GPT-5.4.
META Superintelligence Lab Presents: ProgramBench: Can SOTA AI Recreate Real Executable Programs(ffmpeg, SQLite, ripgrep) From Scratch Without The Internet?
Meta's Superintelligence Lab introduces ProgramBench, a benchmark evaluating whether state-of-the-art AI models can recreate real executable programs like ffmpeg and SQLite from scratch without internet access.
@_philschmid: https://x.com/_philschmid/status/2081744861829414977
EvoCode-Bench is a multi-turn coding benchmark with 26 tasks across 5 domains, designed to evaluate AI agents on evolving specifications and cumulative testing in a persistent workspace, revealing that single-turn scores dramatically overstate reliability.
@OkhayIea: Everyone's racing to build "AI scientists." So we asked a blunt question: Can today's best coding agents beat the publi…
Introduces NatureBench, a cross-disciplinary benchmark of 90 tasks from Nature papers to test AI coding agents, finding the best agent (Claude Opus 4.7) surpasses SOTA on only 17.8% of tasks and often succeeds by reducing science to supervised ML rather than genuine discovery.