Reproduce “Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luxu Liang, Xiang Li
S2D detects generated text by adding a learned vector to intermediate activations of a frozen observer LLM. It scores the resulting representations against class directions using a calibrated threshold. In DetectRL, with a Llama-3.1-8B observer trained on ChatGPT text and tested on Google-PaLM text, it reports 91.93% TPR at 1% FPR, versus RepreGuard’s 81.40%; paraphrasing lowers its TPR to 23.60%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.