Measuring, Localizing, and Ablating Alignment Signatures in LLMs

2026-05-28

AI Summary

This work addresses the pervasive yet poorly understood “AI style” exhibited by large language models after alignment, whose relationship to post-training procedures and internal representations remains unclear. The authors propose PASTA, a training-free decoding-time method that identifies and ablates internal activation directions associated with alignment by comparing generation differences between aligned and base models given identical prefixes. PASTA enables, for the first time, the measurement, localization, and causal intervention of AI stylistic traits introduced during post-training. Experiments across 11 aligned models and 6 detectors demonstrate that PASTA substantially reduces AI detectability with strong generalization across detectors, while preserving text relevance and coherence and even enhancing stylistic diversity.

Abstract

Aligned language models often exhibit a recognizable AI-like style, yet its connection to post-training and internal representations remains poorly understood. In this work, we study whether post-training introduces or amplifies AI-like stylistic regularities and whether these regularities have a localized internal signature. To this end, we compare human text, base-model generations, and aligned-model generations under matched human-source prefixes. Aligned generations show lower human-corpus affinity and higher AI-detection rates than base generations, suggesting that post-training shifts generated text away from human-corpus style and toward detector-visible AI-like text. We then introduce PASTA (Post-training Alignment Signature Targeted Ablation), a training-free method that estimates a post-training alignment signature from aligned-base residual contrasts and ablates the corresponding direction during decoding. Across 11 aligned models and 6 AI detectors, PASTA lowers the detection rate for most aligned models; this effect transfers well across detectors and is not reproduced by random directions. Qualitative analysis suggests that PASTA generations remain relevant and coherent while exhibiting greater stylistic variation. Together, these results show that AI-like stylistic effects of post-training can be measured, localized, and causally tested through activation ablation.

Problem

• alignment
• post-training
• AI-like style
• stylistic regularities
• internal representations

Innovation

• alignment signature
• activation ablation
• AI detection
• post-training
• stylistic regularities