ProgramBench, a new benchmark by researchers, evaluates whether large language models (LLMs) can reconstruct software from scratch. Unlike existing tests focused on isolated fixes, ProgramBench requires models to architect and implement entire codebases matching reference executables. The study includes 200 tasks spanning CLI tools and complex systems like FFmpeg, SQLite, and the PHP interpreter. Results show no model fully resolves any task, with the best passing 95% of tests on just 3% of cases. Models default to monolithic, single-file implementations, diverging from human-written modular code.

The benchmark uses agent-driven fuzzing to generate end-to-end behavioral tests without prescribing implementation structure. This approach forces models to make high-level architectural decisions, such as choosing frameworks or optimizing workflows. However, most LLMs struggle with scalability, often failing to replicate complex interactions in large projects. Human-written code typically employs layered architectures, a contrast starkly highlighted by the models' flat, all-in-one solutions.

ProgramBench’s findings challenge assumptions about AI’s readiness for real-world software engineering. While models excel at generating syntactically correct code, they lack the strategic decision-making required for maintainable, scalable systems. The study underscores the gap between narrow AI capabilities and holistic software development, suggesting future research must address architectural reasoning and long-term codebase management.

Key takeaway: ProgramBench reveals critical limitations in current LLMs for full-stack software development, emphasizing the need for advancements in high-level planning and modular design.