Enhancing LLM Steering through Sparse Autoencoder-Based Vector Refinement

2025-09-28

AI Summary

In few-shot settings, large language model (LLM) steering vectors are highly susceptible to irrelevant noise, limiting their controllability and efficacy. To address this, we propose a steering vector refinement framework based on sparse autoencoders (SAEs). This work is the first to apply SAEs to semantic parsing of steering vectors, leveraging sparse representation learning to automatically disentangle task-relevant features—thereby suppressing noise and compensating for semantic gaps. We further refine and enhance the vectors via semantic similarity-based reconstruction. Experiments demonstrate that our method significantly outperforms existing steering baselines under extreme data scarcity (e.g., 10–50 samples), achieving performance comparable to supervised fine-tuning. It thus effectively alleviates the controllability bottleneck in LLMs induced by limited supervision, enabling robust, interpretable, and data-efficient steering.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Steering has emerged as a promising approach in controlling large language models (LLMs) without modifying model parameters. However, most existing steering methods rely on large-scale datasets to learn clear behavioral information, which limits their applicability in many real-world scenarios. The steering vectors extracted from small dataset often contain task-irrelevant noising features, which degrades their effectiveness. To refine the steering vectors learned from limited data, we introduce Refinement of Steering Vector via Sparse Autoencoder (SAE-RSV) that leverages SAEs to semantically denoise and augment the steering vectors. In our framework, we first remove task-irrelevant features according to their semantics provided by SAEs, and then enrich task-relevant features missing from the small dataset through their semantic similarity to the identified relevant features. Extensive experiments demonstrate that the proposed SAE-RSV substantially outperforms all the baseline methods including supervised fine-tuning. Our findings show that effective steering vector can be constructed from limited training data by refining the original steering vector through SAEs.

Problem

• Refining steering vectors from limited datasets
• Removing task-irrelevant noise using sparse autoencoders
• Enhancing LLM control without parameter modification

Innovation

• Refining steering vectors via sparse autoencoder denoising
• Removing task-irrelevant features using SAE semantics
• Enhancing task-relevant features through semantic similarity augmentation