The Trilemma of Truth in Large Language Models

The study presents sAwMIL (Sparse-Aware Multiple-Instance Learning), a multiclass probing framework designed to assess the veracity of knowledge encoded in large language models (LLMs). Evaluated across 16 open-source LLMs with three newly curated datasets, the framework reveals limitations of existing probing methods and identifies asymmetries in how truth and falsehood are represented in LLMs. This research is significant for practitioners as it provides insights into the reliability of LLM outputs and suggests a new approach to understanding their internal knowledge representation.