AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
This work addresses the need for standardized evaluation in interpretability research for LLM steering, though it is incremental as it builds on existing methods and benchmarks.
The authors tackled the problem of comparing representation-based steering methods for language models by introducing AxBench, a large-scale benchmark, and found that prompting outperforms all existing methods, followed by finetuning, while sparse autoencoders (SAEs) were not competitive.
Fine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce AxBench, a large-scale benchmark for steering and concept detection, and report experiments on Gemma-2-2B and 9B. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. We introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; ReFT-r1), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. Along with AxBench, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.