CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features
This addresses the challenge of automated steering for LLMs in tasks like QA and safety, offering a scalable method that reduces reliance on contrastive datasets.
The paper tackles the problem of effectively steering large language models using sparse autoencoder features by proposing CorrSteer, which selects features by correlating sample correctness with activations at inference time, achieving improvements such as +3.3% on MMLU and +27.2% on HarmBench.
Sparse Autoencoders (SAEs) can extract interpretable features from large language models (LLMs) without supervision. However, their effectiveness in downstream steering tasks is limited by the requirement for contrastive datasets or large activation storage. To address these limitations, we propose CorrSteer, which selects features by correlating sample correctness with SAE activations from generated tokens at inference time. This approach uses only inference-time activations to extract more relevant features, thereby reducing spurious correlations. It also obtains steering coefficients from average activations, automating the entire pipeline. Our method shows improved task performance on QA, bias mitigation, jailbreaking prevention, and reasoning benchmarks on Gemma-2 2B and LLaMA-3.1 8B, notably achieving a +3.3% improvement in MMLU performance with 4000 samples and a +27.2% improvement in HarmBench with only 108 samples. Selected features demonstrate semantically meaningful patterns aligned with each task's requirements, revealing the underlying capabilities that drive performance. Our work establishes correlation-based selection as an effective and scalable approach for automated SAE steering across language model applications.