Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines

2026-05-29

AI Summary

Sparse Autoencoders (SAEs) are often considered ineffective for steering large language models (LLMs). This work reevaluates their steering capability on the AxBench benchmark by introducing an interpretability-driven, supervised feature selection and labeling pipeline. The study demonstrates that the selected SAE features exhibit strong causal influence and enable effective model steering even at low ℓ⁰ sparsity levels, achieving performance comparable to LoRA baselines. These findings challenge the prevailing assumption that high sparsity is essential for successful steering and uncover the underappreciated potential of SAEs in controllable interventions within LLMs.

Abstract

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).

Problem

• Sparse Autoencoders
• model steering
• Large Language Models
• AxBench
• interpretability

Innovation

• Sparse Autoencoders
• Model Steering
• Interpretability
• Supervised Feature Selection
• AxBench