SpecMind: Cognitively Inspired, Interactive Multi-Turn Framework for Postcondition Inference
This addresses the challenge of manual specification writing for programmers, offering an incremental enhancement over existing LLM-based methods.
The paper tackles the problem of generating accurate program postconditions by proposing SpecMind, a multi-turn prompting framework that uses feedback to iteratively refine outputs, resulting in significant improvements in accuracy and completeness over state-of-the-art methods.
Specifications are vital for ensuring program correctness, yet writing them manually remains challenging and time-intensive. Recent large language model (LLM)-based methods have shown successes in generating specifications such as postconditions, but existing single-pass prompting often yields inaccurate results. In this paper, we present SpecMind, a novel framework for postcondition generation that treats LLMs as interactive and exploratory reasoners rather than one-shot generators. SpecMind employs feedback-driven multi-turn prompting approaches, enabling the model to iteratively refine candidate postconditions by incorporating implicit and explicit correctness feedback, while autonomously deciding when to stop. This process fosters deeper code comprehension and improves alignment with true program behavior via exploratory attempts. Our empirical evaluation shows that SpecMind significantly outperforms state-of-the-art approaches in both accuracy and completeness of generated postconditions.