Measuring, Localizing, and Ablating Alignment Signatures in LLMs

In three linesStudy of stylistic signatures introduced by LLM alignment. Researchers show post-training creates a detectable AI-like style. They propose PASTA, a training-free method that localizes and ablates this signature during decoding, reducing detection rates across 11 aligned models and 6 AI detectors.

Summary generated by Claude — human-verified