P^3: A New LLM Workflow for Verified Code Generation
A recent paper published on arXiv (2608.09277) introduces P^3, a novel workflow designed for large language models (LLMs) that integrates the planning of a program alongside its correctness proof. Traditionally, the process of generating verified code follows a linear path: a program is created first, followed by efforts to validate it against a formal specification. This method can often be both inefficient and ineffective, as programs developed without considering their proofs may contain subtle errors or be challenging to verify, resulting in fragile repair cycles. Drawing from Dijkstra's principle that program development and correctness arguments should progress concurrently, P^3 initially formulates a combined program-and-proof plan from the specification, subsequently detailing the implementation and proof framework. This research, authored by a team of researchers, aims to enhance the reliability and efficiency of verified code generation in AI-driven software development.
Key facts
- Paper arXiv:2608.09277 proposes P^3 workflow.
- P^3 stands for Joint Program-and-Proof Planning.
- Traditional verified code generation is sequential: program then proof.
- Sequential pipeline can be ineffective and inefficient.
- Programs generated without proof anticipation can be subtly incorrect or hard to verify.
- P^3 is an LLM-based agentic workflow.
- P^3 derives a unified program-and-proof plan from the specification.
- Inspired by Dijkstra's view that program and correctness argument should be developed together.
Entities
—