Layer-wise Probing Analysis of Belief Encoding in LLMs

Inspired by Professor Tan’s inquiry, “Do LLMs differentiate epistemic belief from non-epistemic belief?”, and the psychological framework established by Vesga et al. (2025), we aim to conduct a layer-wise probing analysis of open-source Large Language Models (LLMs)—specifically OpenAI’s GPT-2 and the Meta Llama series. Our objective is to identify which internal layers encode belief types and determine whether these representations reflect deep semantic understanding or merely surface-level linguistic patterns. Future work may include application to hallucination detection.