arXiv Machine Learning
Measuring, Localizing, and Ablating Alignment Signatures in LLMs
Read the original on arXiv Machine Learning →The Flow has not summarised this story yet — read it at arXiv Machine Learning.
Related stories
arXiv Machine Learning
Oct 23
Measuring, Localizing, and Reducing the AI-Like Writing Style of Post-Trained LLMs
The paper investigates how post‑training of language models introduces or amplifies an AI‑like writing style and whether this style has a localized internal signature. By comparing human text, base‑model outputs, and aligned‑model outputs with matched human‑source prefixes, the authors find that aligned generations are less similar to human corpora and more easily detected as AI‑like. They propose PASTA, a training‑free method that estimates a post‑training alignment signature from residual contrasts and ablates the corresponding direction during decoding, reducing detection rates across multiple models and detectors while preserving relevance and coherence.
arXiv Machine Learning
Sep 18
Sampling Reveals Style: Unsupervised, Training-Free Discovery of Prompt-Conditional Stylistic Axes in LLM Activations
The paper introduces a training‑free method for uncovering prompt‑conditional stylistic axes in large language models (LLMs). By repeatedly sampling completions of a single prompt at high temperature and applying Principal Component Analysis (PCA) to the pooled hidden activations, the authors automatically label the resulting axes using the extreme (pole) generations. Validation against 245 human‑elicited stylistic annotations shows that, for the Qwen‑3.5‑4B‑Instruct model, the top two axes align with human dimensions with 72.8% precision and 43.6% macro‑recall, and 75.6% of validity ratings confirm the axes’ polar generations, while other models exhibit varying degrees of discoverability.
arXiv Computation and Language
Aug 31
AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not
The study demonstrates that text produced by large language models (LLMs) leaves a distinct stylometric footprint—primarily increased entropy and lexical diversity—across multiple models and domains. In contrast, AI editing of human text does not replicate this footprint; edited texts show only modest lexical diversity gains and reduced entropy, with lexical density emerging as the key distinguishing feature. Consequently, stylometric analysis can differentiate AI-generated from AI-edited content, but is less effective at distinguishing either from purely human writing.
arXiv AI
Jun 10
Unsupervised Style Representation Learning for AI-Text Detection via Paraphrase Inversion
arXiv:2606. 10099v1 Announce Type: cross Abstract: The rapid development of large language models (LLMs) has raised concerns about misuse such as plagiarism, misinformation, and automated influence operations, motivating the need for robust detectors.
arXiv AI
Jun 19
How Do Instructions Shape Speech? Cross-Attention Attribution for Style-Captioned Text-to-Speech
arXiv:2606. 20532v1 Announce Type: new Abstract: Style-captioned text-to-speech systems use natural language to control voice characteristics, but how individual words influence acoustic output remains unclear.
arXiv AI
Jul 7
Telescope: Improving Zero Shot Detection of LLM Generated Content By Measuring Token Repetition Probability
arXiv:2607. 04061v1 Announce Type: cross Abstract: Distinguishing Large Language Model (LLM) generated text from human writing is a critical and difficult challenge.