DigestAI news desk

AI news, digested. Every story with its sources, every 30 minutes.

Researchupdated

Study finds PCA can detect stylistic axes in LLM activations without training

Researchers from arXiv cs.CL present a training-free method to identify stylistic dimensions in large language models (LLMs) by analyzing hidden activations. The approach involves sampling completions at elevated temperature and applying Principal Component Analysis (PCA) to the pooled data, automatically labeling the resulting axes based on the generated text poles.

1 source primary source

Key points

  • Training-free PCA on LLM activations detects stylistic axes with 72.8% precision on Qwen-3.5-4B-Instruct.
  • Method validated against 245 human annotations, showing 75.6% accuracy in judging polar generations.
  • DeepSeek-7B-Chat showed only 35.3% precision, indicating stylistic structure is model-dependent.

The study validates these discovered axes against 245 human-elicited stylistic annotations. On the Qwen-3.5-4B-Instruct model, the top two axes matched human-requested dimensions with 72.8% precision and 43.6% macro-recall. Additionally, 75.6% of validity ratings judged the polar generations accurate to their labels, with 90.9% adjacent inter-annotator agreement.

Results show strong model dependency. While Qwen models and Llama-3.2-3B exposed human-salient axes, DeepSeek-7B-Chat dropped to 35.3% precision, with its leading components dominated by structural rather than stylistic variance. The authors conclude that simple PCA over decoding variance is an effective, low-cost probe for understanding stylistic structure in LLM representations and highlighting cross-model differences.

Read the original atarXiv cs.CL · by Ajit Mallavarapu, Ziwei Gu primary sourceOpen source ↗
Topics · follow one to build your own front page
Qwen-3.5-4B-InstructLlama-3.2-3BDeepSeek-7B-Chat

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.

Comments

via GitHub Discussions

More in Research

All →

Related stories