Type: url
Original: https://arxiv.org/pdf/2401.08500
Status: generated
Summary
The full text of a CodiumAI preprint that proposes AlphaCodium, a test-based, multi-stage, code-oriented iterative flow for LLM code generation: a natural-language pre-processing phase (problem reflection, public-test reasoning, candidate solutions, AI-generated tests) followed by iterative run-and-fix cycles against public and AI-generated tests. On the CodeContests competitive-programming dataset it reports that GPT-4 pass@5 on the validation set rose from 19% with a single well-designed direct prompt to 44%, and it describes code-oriented design practices such as YAML structured output, modular code, double validation and test anchors.