Parsel

Parsel is an open-source Python-based framework developed by Eric Zelikman and collaborators for functional decomposition of programmatic tasks into natural-language “Parsel” descriptions, which are then automatically implemented using code LLMs like Codex or GPT‑4. By breaking programs into a hierarchy of modular functions—with natural-language descriptions and unit tests—it explores multiple implementations for each module and combines them to satisfy constraints. This approach enhances correctness and scalability for algorithmic reasoning, improving pass rates on tasks like HumanEval (from 67 % to ~85 %) and APPS by over 75 % compared to direct LLM program synthesis . Parsel is designed to work across domains—from algorithmic problem-solving and program synthesis to robotic planning in simulated environments like VirtualHome—by leveraging decompositional reasoning. It supports automatic test generation and strongly-connected component analysis to handle recursion and dependency structure. It is implemented in Python, and users define task decomposition either manually or via prompting LLMs, after which Parsel orchestrates LLM-based code generation, test-driven validation, and composition into full programs .