WybeCoder: Verified Imperative Code Generation

A research team from Meta AI and the University of Cambridge has introduced WybeCoder, a novel agentic framework that tackles the challenge of verified imperative code generation. The system addresses a significant gap where large language models have advanced code generation but not formal software verification. WybeCoder implements a "prove-as-you-generate" paradigm where code, program invariants, and formal proofs co-evolve simultaneously, building on recent work that combines automatic verification condition generation with SMT solvers and interactive theorem proving in Lean.

To enable rigorous evaluation, the researchers translated two established benchmarks for functional verification—Verina and Clever—into equivalent imperative code specifications. On complex algorithms like Heapsort, WybeCoder demonstrated consistent performance improvements by scaling its approach, synthesizing dozens of valid invariants and dispatching numerous verification subgoals. This resulted in the generation of hundreds of lines of formally verified imperative code, overcoming plateaus reported in previous verification attempts. The system's best configuration achieved a 74% success rate on Verina tasks and 62% on Clever tasks using moderate computational budgets, significantly surpassing previous evaluations in this domain.

The framework's architecture enables systematic exploration of the verification space, allowing it to navigate through complex proof obligations that typically stall automated provers. By integrating multiple verification techniques and scaling them through an agentic approach, WybeCoder paves the way for automated construction of large-scale datasets of verified imperative code. This advancement could accelerate the development of more reliable software systems and provide valuable training data for future AI-assisted programming tools focused on correctness.