Tacl Repreguard

One paper accepted by Transactions of the Association for Computational Linguistics (TACL) as first author: RepreGuard: Detecting LLM-Generated Text by Revealing Hidden Representation Patterns. We propose RepreGuard, a low-overhead, highly generalizable, and interpretable detector based on the observation that there are significant differences in neural activation patterns when LLMs process LLM-generated text versus human-written text.