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

Researchers present LLM+ASP, a framework combining large language models with Answer Set Programming to enable nonmonotonic reasoning without task-specific engineering. The system uses automated self-correction loops where an ASP solver provides structured feedback, demonstrating significant performance improvements over monotonic logic approaches across diverse reasoning benchmarks.

The LLM+ASP framework addresses a critical limitation in current AI systems: the inability to handle defeasible reasoning, where conclusions can be revised based on new information. Traditional neuro-symbolic approaches couple LLMs with monotonic logics like SMT solvers, which enforce logical consistency but cannot represent the nuanced exception-handling that characterizes human reasoning. This research demonstrates that Answer Set Programming's stable model semantics naturally express default rules and their exceptions, making it superior for complex reasoning tasks where circumstances change.

The breakthrough lies in the self-correction mechanism. Rather than requiring domain experts to author knowledge modules or craft specialized prompts, the framework iteratively refines outputs based on solver feedback. Testing across six diverse benchmarks reveals that this automated correction process is the primary performance driver, effectively eliminating the need for handcrafted domain knowledge. Additionally, the research identifies a counterintuitive phenomenon: compact reference guides outperform verbose documentation, suggesting that excessive context actually impairs constraint adherence—a finding with implications for prompt engineering practices industry-wide.

For the AI development community, this work reduces barriers to deploying symbolic reasoning systems. Eliminating per-task engineering means organizations can apply LLM+ASP across diverse problem classes without specialized expertise. The nonmonotonic reasoning capability enables more human-like decision-making in complex domains like legal reasoning, medical diagnosis, and planning under uncertainty.

Looking forward, the challenge lies in scaling this approach to real-world complexity while managing computational costs. The research addresses LLM limitations but doesn't fully resolve the high computational expenses mentioned in the abstract. Future work should explore how this framework performs on production-scale problems and whether the efficiency gains justify implementation complexity.

• →LLM+ASP enables nonmonotonic reasoning without per-task engineering or domain-specific prompts, unlike prior neuro-symbolic approaches.
• →Automated self-correction loops driven by ASP solver feedback prove more effective than manually authored knowledge modules.
• →Stable model semantics outperform monotonic logics (SMT) for tasks requiring default rules and exception handling.
• →Compact reference guides improve constraint adherence better than verbose documentation, revealing a 'context rot' phenomenon.
• →The framework operates uniformly across diverse reasoning tasks, reducing implementation barriers for deploying symbolic AI systems.