Hallucination Detection with the Internal Layers of LLMs

Abstract

Large Language Models (LLMs) have succeeded in a variety of natural language processing tasks [Zha+25]. However, they have notable limitations. LLMs tend to generate hallucinations, a seemingly plausible yet factually unsupported output [Hua+24], which have serious real-world consequences [Kay23; Rum+24]. Recent work has shown that probing-based classifiers that utilize LLMs' internal representations can detect hallucinations [AM23; Bei+24; Bur+24; DYT24; Ji+24; SMZ24; Su+24]. This approach, since it does not involve model training, can enhance reliability without significantly increasing computational costs. Building upon this approach, this thesis proposed novel methods for hallucination detection using LLM internal representations and evaluated them across three benchmarks: TruthfulQA, HaluEval, and ReFact. Specifically, a new architecture that dynamically weights and combines internal LLM layers was developed to improve hallucination detection performance. Throughout extensive experiments, two key findings were obtained: First, the proposed approach was shown to achieve superior performance compared to traditional probing methods, though generalization across benchmarks and LLMs remains challenging. Second, these generalization limitations were demonstrated to be mitigated through cross-benchmark training and parameter freezing. While not consistently improving, both techniques yielded better performance on individual benchmarks and reduced performance degradation when transferred to other benchmarks. These findings open new avenues for improving LLM reliability through internal representation analysis.