Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselines

Steering LLMs? Actually, Sparse Autoencoders can outper- form simple baselines Mikkel Godsk Jørgensen mikkel.godsk.research@gmail.com DTU Compute Technical University of Denmark Lars Kai Hansen lkai@dtu.dk DTU Compute Technical University of Denmark Abstract Sparse Autoencoders (SAEs) have been seen as a promising avenue for exploring the inter- nals 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 sug- gests 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). Our code is available at https://github.com/MikkelGodsk/SAE-labelling. 1 Introduction Sparse Autoencoders have recently attracted a wave of attention (e.g. Cunningham et al., 2023; Makelov et al., 2024; Templeton et al., 2024; Gao et al., 2024; Rajamanoharan et al., 2024), offering an unsupervised interpretability model based on the Linear Representation Hypothesis (LRH) and the superposition hypoth- esis. The LRH suggests that many concepts may be stored in linear subspaces, as discussed in e.g. Alain & Bengio (2018); Kim et al. (2018); Elhage et al. (2022); Park et al. (2023), although somewhat contested in e.g. Crabbé & van der Schaar (2022). Since LLMs can represent a large number of concepts in relatively low-dimensional latent spaces, the neurons must be polysemantic as a consequence, i.e. they become ac- tive across multiple different semantics, as discussed in e.g. Bolukbasi et al. (2021); Makelov et al. (2024); Templeton et al. (2024); Fereidouni et al. (2025); Cunningham et al. (2023). The superposition hypothesis further suggests that linear concept representations are densely packed and not fully orthogonal (Elhage et al., 2022). Within this framework, Sparse Autoencoders offer an appealing interpretability tool, capable of discovering a subset of linear concept representations, referred to as features, which are thought to exhibit a high degree of monosemanticity (e.g. Bricken et al., 2023; Templeton et al., 2024). In Fereidouni et al. (2025), the authors quantify the monosemantic property of SAEs in comparison to individual neurons in a network. Their findings suggest that, indeed, SAE features are more monosemantic than the individual neurons of language models. While the monosemantic response is interesting for interpretability, features have also shown some ability to control the model’s text generation (e.g. Templeton et al., 2024; Arad et al., 2025; He et al., 2025; Wang et al., 2025), although the effectiveness is disputed most notably in Wu et al. (2025). In fact, Wu et al. (2025) concludes that “Even simple baselines outperform Sparse Autoencoders” for steering, and finds that 1 arXiv:2605.31183v1 [cs.CL] 29 May 2026

simply prompting the model is much more effective than other approaches, such as intervention via SAEs (e.g. Templeton et al., 2024) and LoRA (Hu et al., 2022). Although this is undeniably important for most practical applications, the prompt baseline can seem unfair since the subject LLM was originally trained precisely to follow instructions. Aside from steering, there have also been several attempts at using SAEs for knowledge unlearning/concept removal in language models (e.g. Farrell et al., 2024; Yamashita et al., 2025; Fereidouni et al., 2025), essentially exploring the negative (removal) dimension of steering. Since Sparse Autoencoders are trained unsupervised, efforts have been made to label their features with the concepts they represent. This has been attempted at scale using LLMs (e.g. Lieberum et al., 2024; Paulo et al., 2024), but to our knowledge, no dedicated efforts have been made to use a supervised approach to labelling. Recent work (Wang et al., 2025) has suggested that high interpretability of a feature tends to have some corre- lation with high utility for model steering, and that this relationship seems most pronounced for JumpReLU SAEs with a low ℓ0 (i.e. high sparsity). This result could pose as a relevant context for interpreting the findings of Wu et al. (2025), in which they base their conclusions on SAEs with high ℓ0, being those that are already annotated in Neuronpedia1. 1.1 Contributions We make the following contributions: 1. We introduce a labelling pipeline for Sparse Autoencoders based on labelled datasets rather than LLMs. 2. We demonstrate that Sparse Autoencoders are substantially more powerful for steering than re- ported in Wu et al. (2025) when the features are labelled via our proposed method. However, the steering performance we demonstrate is still not fully competitive with prompting on the AxBench benchmark. 3. We find indications that both low- and high-sparsity Sparse Autoencoders in Gemma Scope can be used for steering. 2 Preliminaries 2.1 Large Language Models Formal definition. Large Language Models define a distribution over text strings, and are typically based on transformers (Vaswani et al., 2023). To transform a string of text into a format suitable for language modeling, the typical approach is to use a tokenizer to split up the text into a string of tokens. Here, the tokens constitute an alphabet Σ. For modeling purposes, we further augment the alphabet by the tokens BOS and EOS (beginning of sequence, end of sequence) to get an augmented alphabet ¯Σ = Σ ∪ {BOS, EOS}. Formally, a sequence of tokens is simply called a string. The language model can then be characterized as a string distribution pLM(s) = ∏ i pSM(si|s<i), s ∈ (BOS · Σ∗ · EOS) , (1) where · denotes string concatenation and pSM(·|s<i) : ¯Σ → [0, 1] is typically fitted by the model. Here pLM is assumed to be a true distribution (i.e., no probability mass is leaked into the set of infinite strings). This is true for models with a softmax-based prediction head. To use the language model generatively, we use a decoding strategy to auto-regressively predict the next token until termination by EOS. The two simplest options are to greedily select arg maxsi pSM(si|s<i) as our next token, or to sample from pSM(si|s<i) at each step (Cotterell et al., 2024). Transformers. LLMs typically build on a modified transformer architecture compared to the one introduced in Vaswani et al. (2023). The main modifications are to use only the decoder, to apply the positional 1https://www.neuronpedia.org/gemma-scope#browse 2

fora: academia, biology, history, law, literature, physics, and sports. Hence, the remaining fora and layer 17 are not included. As elaborated in section 3.1, K and the output score threshold were chosen heuristically in the early experimentation in this setup. We find that pipeline performance seems particularly sensitive to the choices of π0 and the support threshold. Since it is infeasible to evaluate it on the full Cartesian product of features × labels, we use a manual iterative approach. Here, we select sensible initial values and evaluate the assigned labels. We then inspect the results, tune the parameters optimizing for concept_rating, and repeat. After a few iterations, we converge at π0 = 10−3 and a support threshold of 50, still being sensible values. Baseline. As baselines, we use those provided in Wu et al. (2025), including a reproduction of the SAE and prompt baselines named NP random. NP random samples 100 features at random and evaluates their official Neuronpedia labels for steering and prompting. Interpretation. Comparisons of steering performances between feature-steering and prompt-steering are shown in figures 2 and 6 (see appendix B). In figure 6, we see that our proposed pipeline is indeed capable of assigning correct labels to the causal effects of features. While we observe some variability between different fora, we mostly observe similar effects when comparing feature-steering to prompt-steering. Figure 2 depicts the overall steering performance and shows a general improvement over those reported in Wu et al. (2025). We observe that multiple fora provide labels that result in feature-steering performance on par with the LoRA reference. We also observe a clear improvement in the prompt-steering performance, possibly due to the subject model’s better understanding of simpler labels. The reference Neuronpedia performance is largely consistent with our results. When considering the performance increase of both feature-steering and prompt-steering from the Neuron- pedia reference, one could be tempted to alternatively suggest better label comprehension by the LLM judge as an explanation. However, when considering the concept_rating performance in figure 6, we now observe a mostly closed gap between feature-steering and prompt-steering compared to the Neuronpedia baseline. Since concept_rating is the only component in AxBench that relies on the feature-label, the alternative explanation does not fully account for the results. Ablation. We perform an ablation study to examine the impacts of the roles of 1) a high sparsity, and 2) the output score. The results are shown in figures 7 and 8. Here we see that using high sparsity SAEs (i.e., low ℓ0) seemingly does not greatly impact the steering performance. We also observe a rather modest performance improvement caused by the output score, in both the low and high sparsity SAEs. To examine the role of the output score, we compare randomly selected features with features selected by our pipeline. In both feature sets, we fetch their original Neuronpedia labels. Here, we observe better feature-steering performance of those selected by the pipeline, even without using the output score. 5 Discussion We find the results of section 4 somewhat surprising for the two reasons: 1) causality generally cannot be taken for granted given a correlation/dependence, yet removing the causal-component of the pipeline (the output-score) does not greatly affect its performance; and 2) the reported feature-steering performance in Wu et al. (2025) was much weaker compared to our setup. Linear Representation Hypothesis. While Wu et al. (2025) reported obtaining rather poor steering per- formance on all methods that linearly steer, we found that SAEs can, surprisingly, steer with a performance close to par with LoRA that uses adapter weights (Hu et al., 2022). This finding demonstrates the power of using linear methods for understanding the internals of language models. AxBench. Since AxBench uses an LLM-judge to rate steered generations, we find it natural to ask whether the LLM-judge could exhibit a bias to favor some steering methods differently than a population of human judges. To investigate this, we conducted a small trial with Danish bachelor students and found indications that they would generally rate some scores more optimistically than GPT-4o-mini, however, we did not find clear evidence of a bias in the trial. We note that the trial had a small sample size, so a more thorough user study might still be relevant. However, while inspecting our pipeline evaluation results, we observed multiple 7