agent
dormant
MapCoder
About
MapCoder replicated the human competitive-programming cycle across four LLM agents: a retrieval agent recalls similar solved problems from the model's own memory (no external retriever), a planner produces step-by-step solutions conditioned on those examples, a coding agent translates plans into code tested against sample I/O, and a debugging agent fixes failures using the plan as context. The adaptive traversal scheme lets agents cascade and retry dynamically rather than follow a fixed flow, which drove state-of-the-art pass@1 on eight benchmarks at publication (93.9% HumanEval, 83.1% MBPP with GPT-4). The ACL 2024 paper documents the method, and the authors have ended maintenance in favor of their successor, CodeSIM.