Parsel, built by Eric Zelikman and collaborators, tackles a specific weakness in how LLMs write code: they tend to do worse on complex, multi-step programs than on small, isolated functions. It addresses that by decomposing a task into a hierarchy of modular functions, each described in natural language with accompanying unit tests, then has a code LLM like Codex or GPT-4 generate and test multiple candidate implementations for each piece before combining them into a full program.
That decompose-generate-test-combine approach improved HumanEval pass rates from 67% to roughly 85% in the original research, and showed similarly large gains on the harder APPS benchmark, suggesting that breaking a problem down structurally helps more than simply prompting a bigger model with the whole task at once. The same approach extended beyond pure coding into robotic planning in simulated environments like VirtualHome.
It's implemented in Python and remains primarily a research artifact rather than a polished developer product, useful for understanding decompositional prompting techniques more than as a tool to install and use day to day.