AutoICE: Automatically Synthesizing Verifiable C Code via LLM-driven Evolution

TL;DR

AutoICE leverages large language models and evolutionary search techniques to automatically synthesize verifiable C code from natural language requirements, improving correctness and reliability.

Contribution

It introduces diverse initialization, collaborative crossover, and self-reflective mutation to enhance verifiable code synthesis, addressing errors and implicit knowledge formalization.

Findings

Achieves 90.36% verification success rate on standard datasets.

Outperforms state-of-the-art methods in code verification accuracy.

Attains 88.33% success rate on developer-friendly dataset.

Abstract

Automatically synthesizing verifiable code from natural language requirements ensures software correctness and reliability while significantly lowering the barrier to adopting the techniques of formal methods. With the rise of large language models (LLMs), long-standing efforts at autoformalization have gained new momentum. However, existing approaches suffer from severe syntactic and semantic errors due to the scarcity of domain-specific pre-training corpora and often fail to formalize implicit knowledge effectively. In this paper, we propose AutoICE, an LLM-driven evolutionary search for synthesizing verifiable C code. It introduces the diverse individual initialization and the collaborative crossover to enable diverse iterative updates, thereby mitigating error propagation inherent in single-agent iterations. Besides, it employs the self-reflective mutation to facilitate the discovery of implicit knowledge. Evaluation results demonstrate the effectiveness of AutoICE: it successfully verifies $90.36$\% of code, outperforming the state-of-the-art (SOTA) approach. Besides, on a developer-friendly dataset variant, AutoICE achieves a $88.33$\% verification success rate, significantly surpassing the $65$\% success rate of the SOTA approach.