講演情報
[4K1-GS-6a-04]Analysis of the Discrepancy Mechanism between Internal Knowledge and Surface Output in LLMs and Falsehood Suppression using Sparse Autoencoders
〇Jifan Zhang1, Tengfei Shao1, Ayako Yamagiwa1, Masayuki Goto1 (1. Waseda University)
キーワード：
Large language model、Sparse autoencoder、Representation interpretability、Falsehood generation、Inference-time intervention
Large Language Models (LLMs) are known to output "imitative falsehoods" in response to inductive inputs, even when possessing correct internal knowledge. TruthfulQA, a benchmark evaluating truthfulness by testing a model's ability to resist such induction, remains a challenging task for LLMs. Existing methods to address this issue often face problems with robustness or increased inference costs. In this study, we propose an inference-time intervention method using Sparse Autoencoders (SAEs) to reduce the discrepancy between a model's internal knowledge and surface output. Specifically, we decompose intermediate layer activations using SAEs to identify feature directions that trigger false outputs or correlate with truthfulness. We then utilize these features during inference to suppress the falsehood generation mechanism. This method requires no retraining and serves as a practical approach to elicit the latent truthfulness of LLMs. We verify the effectiveness of our proposal by applying it to the TruthfulQA benchmark.