Overview
Large language models (LLMs) encode rich stylistic structure in their hidden activations, but discovering which stylistic dimensions are salient for a given prompt typically requires supervised contrastive data. This paper presents a training-free, prompt-conditional alternative for probing that structure.
Method
The proposed approach is deliberately simple:
- Repeated sampling — Generate multiple completions of a single prompt at elevated temperature.
- Dimensionality reduction — Apply Principal Component Analysis (PCA) to the pooled hidden activations.
- Automatic labeling — Label the resulting axes automatically from the pole generations.
- 72.8% precision
- 43.6% macro-recall
- 75.6% of validity ratings judging the axes' polar generations accurate to their labels
- 90.9% adjacent inter-annotator agreement
- Both Qwen models and Llama-3.2-3B expose human-salient axes.
- DeepSeek-7B-Chat drops to 35.3% precision, with its leading components dominated by structural rather than stylistic variance.
No fine-tuning, contrastive pairs, or supervision is required.
Validation
The discovered axes were validated against 245 human-elicited stylistic annotations in a two-phase study.
On the strongest model tested — Qwen-3.5-4B-Instruct — the top two axes match spontaneously requested human dimensions with:
Model Dependence
Discoverability is strongly model-dependent:
Conclusion
Simple PCA over a model's own decoding variance is an effective, low-cost probe of stylistic structure in LLM representations. Beyond efficiency, the method also exposes sharp cross-model differences in how that stylistic structure is organized — a finding relevant to interpretability and model comparison work in 2026, as LLM activation analysis continues to mature into a standard diagnostic toolkit.
via ArXiv CL+LG