Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines
This work provides a partial rebuttal to previous findings for researchers exploring the internal mechanisms and steering capabilities of LLMs, suggesting that SAEs are more effective than previously thought.
This paper re-evaluates Sparse Autoencoders (SAEs) for steering Large Language Models (LLMs) on the AxBench benchmark, finding that with a supervised feature selection and labeling pipeline, SAEs can perform nearly as well as LoRA. The pipeline also selects surprisingly causal features for their identified labels, and suggests that high sparsity may not be crucial for interpretability-based steering.
Sparse Autoencoders (SAEs) have been seen as a promising avenue for exploring the internals of Large Language Models (LLMs) and for steering model output generation. When AxBench - a model steering benchmark - was introduced in Wu et al. (2025), SAEs did not seem to live up to their original hype due to poor steering performance relative to a set of simple baselines. This work serves as a partial rebuttal for Sparse Autoencoders and suggests that the results of Wu et al. (2025) did not do them full justice. We find that Sparse Autoencoders can, in fact, perform close to on par with the reference LoRA performance on the AxBench benchmark, when features are selected and labelled with our supervised pipeline. We also find that our pipeline selects features that are surprisingly causal of their identified labels when using only its interpretability-based components. Lastly, we present evidence that high sparsity (low l0) may not be crucial for successful steering based on interpretability, which is in contrast to the earlier findings in Wang et al. (2025).