Sparse autoencoder (SAE)-based activation steering is a lightweight approach for modifying the behavior of large language models (LLMs) by adding small vectors to hidden states during inference. However, designing effective steering interventions remains challenging. In this work, we develop an information-theoretic framework for analyzing SAE activations through empirical feature-firing distributions. The central object of the framework is the Jensen--Shannon divergence (JSD) between such firing distributions, which quantifies how far a generated or intervened state deviates from a dataset-induced target distribution. Based on an analytic divergence response for virtually shifted states, we construct steering vectors toward the target firing distribution. The same divergence can also be used to monitor deviation from another empirical distribution, such as a general valid-language distribution. We demonstrate the effectiveness of the steering in a question-mode generation task study. We further use teacher-forced pulse experiments to show that local interventions induce structured downstream JSD responses in later layers, and apply the framework to a safety-related target--contrast response setting for feature analysis.
Figures & tables
Figure 1: Overview of DESSIRT . For each dataset, we obtain an SAE feature firing-probability vector from residual-stream activations. We introduce information-theoretic divergence to measure the distinguishability between two datasets. Our objective for constructing a steering vector is to minimize the divergence between the target distribution and anchor distribution under a virtual shift parameterized by a scalar coefficient α . Furthermore, we introduce the cost to suppress the deviation from the general text distribution to preserve fluency.
Figure 2: Mode congruence score as a function of the steering coefficient α for the seven question modes. We connect the points with lines only when the fluency/format score is above 4.0.
Figure 3: Mode congruence score as a function of fluency/format score for the seven question modes.
Figure 4: SAE-based JSD obtained in the teacher-forced experiment. The left five subfigures correspond to the JSD for layers 16,20,24,28, 31 in the current token-generation step, and the rightmost subfigure correspond to that for layer 16 in the next token-generation step. In each subfigure, the blue/green (orange/red) line indicates JSD between the output distribution and the target (all questions) distribution. The green and red lines are JSDs for the teacher-then-free (TTF) condition.
M0
M1
M2
M3
M4
M5
M6
DESSIRT-2, L16
0.700 (4.62)
0.440 (4.72)
0.495 (4.63)
0.970 (4.57)
0.995 (4.54)
0.850 (4.72)
0.895 (4.60)
SDCV, L16
0.565 (4.69)
0.385 (4.70)
0.725 (4.59)
0.900 (4.99)
0.935 (4.58)
0.490 (4.58)
0.570 (4.79)
Prompt
0.895 (4.74)
0.855 (4.77)
0.375 (4.89)
0.950 (4.79)
0.760 (4.16)
0.545 (4.08)
0.770 (4.08)
Table 1: Congruence scores on (a) Llama-3.1-8B-Instruct and (b) Gemma-2-9B-IT. The values in parentheses denote the mean format score.
S0
S1
S2
S3
S4
S5
S6
S7
S8
S9
S10
fluency
no-steering
4.45
4.67
4.47
4.83
4.51
4.95
4.07
4.81
5.00
4.38
4.58
4.76
DESSIRT-1
4.84
4.93
4.86
4.97
4.89
4.95
4.11
4.93
4.99
4.85
4.86
4.99
DESSIRT-2, budget=10
4.79
4.91
4.79
4.92
4.85
5.00
4.07
4.86
4.95
4.70
4.82
4.92
DESSIRT-2, budget=30
4.70
4.85
4.73
4.93
4.72
5.00
4.09
4.85
5.00
4.65
4.82
4.95
SAE-CAA
4.69
4.84
4.65
4.93
4.76
4.94
4.03
4.93
5.00
4.55
4.76
4.94
SDCV
4.63
4.74
4.71
4.87
4.72
4.95
3.91
4.81
5.00
4.60
4.72
4.88
Table 2: LLM-as-a-judge evaluation of eleven safety metrics and fluency score for model outputs. Except for the no-steering case, the number of selected features is ten and the steering coefficient is set to α=2 . The safety metrics are: (S0) non deceptive content, (S1) no harassing language, (S2) harmlessness, (S3) no hate speech, (S4) legal compliance, (S5) privacy preservation, (S6) no sensitive disclosure, (S7) no self harm content, (S8) sexual content safety, (S9) ethical compliance, and (S10) no violent content. We use the score range [0,5].
Figure 5: Feature distributions in the (pA,pB) plane, where pA and pB are firing probabilities on datasets A and B , respectively. (a) DESSIRT-1. (b) SDCV-based selection. Colors indicate the rank in the selection order (1–10, 11–100, 101–300, 301–1000, 1001+).
toxic data samples selected according to the perplexity score
Table 3: Datasets used in the SAE training and the number of extracted vectors from each dataset. The main SAE uses all three corpora listed here.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Coefficient-resolved detected-mode heatmaps for (a) the DESSIRT with 30 features and budget=30 and (b) the SDCV baseline with 30 selected features. Each panel is conditioned on a target mode. The horizontal axis shows the steering coefficient α , the vertical axis shows the detected mode, and the color indicates the number of samples. Bright mass on the target-mode row corresponds to successful steering.
Original Dataset
#samples
subsets
WildGuardMix
1900
not adversarial but harmful data samples are selected (samples used in the SAE training are excluded)
AILuminate Demo
1200
all
AttaQ
1402
all
OR-bench
500
sampled from or-bench-toxic
Appendix
Table 4: Prompt datasets used in generating datasets A and B .
Rank
Type (rough)
Source
Examples of high-activation text fragments
#1
SC
A
“. It’s important”
“with your inquiry,”
B
“. Instead, I”
“.\n\n- I’m”
#2
SC
A
“To further explore strategies”
“ additional insights tailored to”
Appendix
Table 6: Representative high-activation fragments for the top ten DESSIRT-1-selected features. Fragments are shown as illustrative examples; each feature also activates on other related phrases that are not listed here. SA =safety-relevant content, RS =response stance, SC =scope control, IN =interpersonal.
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).
Latent steering exploits internal representations of Large Language Models (LLMs) to guide generation, yet interventions on dense states can entangle distinct semantic features. In this paper, we investigate attention query activations as a high-fidelity site for precise control, hypothesizing that manipulating the attention mechanism itself offers sharper steerability than general state interventions. We introduce Prototype-Based Sparse Steering, a framework that applies Sparse Autoencoders (SAEs) specifically to query activations, to decompose them into interpretable features, then apply gradient-based optimization during inference to align the sparse representation with class prototypes of target behaviors. To validate this architectural insight, we first analyze the mechanism in Textualized Gridworld, a controlled environment for verifiable planning constraints. We demonstrate that optimizing sparse query features enables effective navigation of rigid planning requirements (i.e., safe vs. short paths), confirming the method's ability to satisfy objective rules. We then demonstrate the framework's versatility by training SAEs on a high-dimensional educational domain, where the framework steers the cognitive complexity of feedback (i.e., Bloom's Taxonomy). Our experiments establish that sparse query representations provide the necessary disentanglement for unified, interpretable control over both logical planning and stylistic nuance.
Sumanta Bhattacharyya, Pedram Rooshenas
Department of Computer Science University of Illinois Chicago
Sparse autoencoders (SAEs) decompose language model activations into interpretable features, but existing methods reveal only which features activate, not which change model outputs when amplified. We introduce Control Reinforcement Learning (CRL), which trains a policy to select SAE features for steering at each token, producing interpretable intervention logs: the learned policy identifies features that change model outputs when amplified. Adaptive Feature Masking encourages diverse feature discovery while preserving singlefeature interpretability. The framework yields new analysis capabilities: branch point tracking locates tokens where feature choice determines output correctness; critic trajectory analysis separates policy limitations from value estimation errors; layer-wise comparison reveals syntactic features in early layers and semantic features in later layers. On Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest, CRL achieves improvements while providing per-token intervention logs. These results establish learned feature steering as a mechanistic interpretability tool that complements static feature analysis with dynamic intervention probes