English

Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering

Computation and Language 2025-07-31 v2 Artificial Intelligence

Abstract

Linear concept vectors effectively steer LLMs, but existing methods suffer from noisy features in diverse datasets that undermine steering robustness. We propose Sparse Autoencoder-Denoised Concept Vectors (SDCV), which selectively keep the most discriminative SAE latents while reconstructing hidden representations. Our key insight is that concept-relevant signals can be explicitly separated from dataset noise by scaling up activations of top-k latents that best differentiate positive and negative samples. Applied to linear probing and difference-in-mean, SDCV consistently improves steering success rates by 4-16\% across six challenging concepts, while maintaining topic relevance.

Keywords

Cite

@article{arxiv.2505.15038,
  title  = {Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering},
  author = {Haiyan Zhao and Xuansheng Wu and Fan Yang and Bo Shen and Ninghao Liu and Mengnan Du},
  journal= {arXiv preprint arXiv:2505.15038},
  year   = {2025}
}

Comments

12 pages, 4 figures, 4 tables

R2 v1 2026-07-01T02:27:08.118Z