Reproduce “Linear Probes Transfer from Base to Distilled Language Models despite Rotation” 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
Marios Tsatsos
The authors test whether base-trained linear probes transfer unchanged to distilled and instruction-tuned successors across four Qwen2.5/Llama-3.1 model families. Despite safety-direction rotations of 17°–35°, observed AUC remained similar on XSTest and an independent evaluation. On OR-Bench, probes flagged 29%–42% of benign-but-alarming prompts at an unvalidated default threshold, supporting reuse only for the tested models and evaluations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.