Hybrid Symbolic AI Outperforms Prompt-Based Methods for Policy Reasoning

A new paper from arXiv introduces a hybrid symbolic approach for generative AI systems to reason over written policies, such as tax rules or airline baggage allowances. The method expresses policies in formal logic and uses language models for fact extraction to ground predicates, while an answer set solver performs reasoning. This separation of extraction and reasoning steps makes responses interpretable, auditable, accurate, and robust under input perturbations. In tests, the approach outperformed policy-as-prompt and policy-as-code methods in most cases, with a ~10x reduction in token usage. The paper, titled 'Policy-as-logic for robust reasoning over rules,' is available on arXiv under the identifier 2608.11905, submitted in the Computer Science > Artificial Intelligence category. The authors argue that structured reasoning and symbolic solvers combined with generative models are valuable for making robust decisions involving objective criteria.