Probabilistic Framework for Commonsense Reasoning in LLMs
A new arXiv preprint (2605.08011) introduces PACS (Probabilistic Abductive CommonSense), a framework that combines large language models with formal logic solvers to handle variability in commonsense beliefs. Unlike prior methods that assume universal agreement, PACS models individual differences by sampling proofs as observations of distinct commonsense assumptions. The algorithm aggregates conclusions to determine whether most people would judge a statement true or false, addressing a key limitation in neurosymbolic reasoning.
Key facts
- arXiv preprint 2605.08011
- Introduces PACS algorithm
- Combines LLMs with formal logic solvers
- Models variability in commonsense beliefs
- Samples proofs as observations of individual beliefs
- Aggregates conclusions across samples
- Addresses assumption of universal commonsense agreement
- Focuses on abductive reasoning
Entities
Institutions
- arXiv