LLMs as ASP Programmers: Self-Correction Enables Task-Agnostic Nonmonotonic Reasoning

The article introduces the "LLM+ASP" framework, which integrates large language models (LLMs) with Answer Set Programming (ASP) to enable task-agnostic nonmonotonic reasoning. This system eliminates the need for manual knowledge engineering and domain-specific prompts, utilizing an automated self-correction loop that enhances performance across diverse reasoning tasks. Evaluations show that this approach significantly outperforms traditional SMT-based methods in handling nonmonotonic tasks, emphasizing the importance of iterative self-correction and the efficiency of compact in-context references over verbose documentation for improving reasoning capabilities in LLMs.