ProgramBench: Can Language Models Rebuild Programs from Scratch?

May 7, 2026 · ai ai-agents coding · Source ↗

TLDR

arXiv paper introduces ProgramBench, finding no LLM can fully reconstruct real programs like FFmpeg or SQLite from docs alone.

ProgramBench tasks agents to re-implement 200 programs using only the reference executable and its documentation, without prescribing code structure.
Evaluation uses agent-driven fuzzing to generate behavioral tests, avoiding implementation-structure bias.
Tasks span compact CLI tools to major projects: FFmpeg, SQLite, PHP interpreter.
Best model passes 95% of tests on only 3% of tasks; no model fully resolves any task across 9 LLMs tested.
Models default to monolithic single-file implementations, diverging sharply from human-written modular codebases.

Discussion is thin and largely skeptical; commenters anticipate benchmark results will be dismissed by practitioners who claim the right agent setup wasn’t used.
One commenter sarcastically extrapolates to a future where models produce machine code directly, skipping compilers entirely, flagged as /s.

@makerofthings: “If it works, the AI is magic. If it doesn’t work, you’re using it wrong” – captures the unfalsifiability critique neatly.