Representation steering methods like difference-in-means (DIM) offer a lightweight approach to controlling LLM behavior, but their effectiveness varies widely across concepts, datasets, and layers, making principled deployment costly without exhaustive hyperparameter sweeps. We build on the discriminability index d′, a compute-cheap geometric measure of how well contrastive training activations separate along the steering direction. Across five behaviors on 3 models, we demonstrate uses beyond its established role as a dataset-level steerability predictor. First, in the multiple-choice question answering (MCQA) setting, we show that d′ reliably identifies performant layers at which to steer, eliminating the need for a layer-wise sweep. Second, we extend this analysis to open-ended multi-token generation, and find that d′ remains successful for layer selection across different settings. Lastly, we test whether the position of a generated token's representation on the DIM line predicts steering success on that individual example. In the MCQA setting, this works for high-d′ datasets but is inconsistent across behaviors, models, and training-token choices in open-ended generation. This highlights that methods developed and conclusions drawn in single-token MCQA settings do not necessarily transfer to open-ended generation.
Figures & tables
Figure 1: Difference-in-means (DIM) steering leverages a set of contrastive training data pairs (red and green points, representing LLM hidden states at some layer ℓ ) to define a linear direction between their centroids ( μ− and μ+ ) that becomes steering vector v . At test-time, this steering vector is added to a test point to move it in the direction of the positive centroid, scaled by some α (blue line; here α=1 ). The best layer ℓ∗ is selected via an expensive grid search over layers to optimize validation steering success. Each datapoint a can be projected onto the line defined by v ; κa Braun (2026) is its scalar distance from centroid μ . d′ Braun et al. (2025) is a summary statistic of the training data’s κa (bottom line; here, indicating generally high separability between classes). We propose (1) using high d′ as a cheap signal to choose which layer to steer at, avoiding the full layer sweep, (2) studying generalization from single-token MCQA to multi-token open-ended generation, and (3) exploring individual steered test points’ κa as a predictor of steering success.
Figure 2: MCQA per-layer test steering accuracy at the best α (top) and training d′ (bottom) for Gemma-2-9B-IT. The vertical line marks the top-1 d′ layer, circles the top-3 d′ layers, and the star the best layer in the full sweep; the dotted line is the unsteered baseline. For most behaviors, the highest- d′ layers lie at or near the best layer. Similar figures for other models are in Section B.1 .
Table 1: MCQA results across five behaviors and three models using αl∗ . Top-1 is steering at the layer with the highest training d′ , top-3 is the layer with the highest validation accuracy among the three highest- d′ layers, and best is the layer with the highest validation accuracy in the full sweep. All selections are made on the validation set; all reported numbers are on the test set. The last row gives the median share of the best achievable gain over baseline recovered by top-1 / top-3, over model × behavior pairs where steering improves on the baseline.
Behavior
Gemma-2-9B-IT
Llama-3.1-8B
Qwen3.5-9B
Mean-pooled Steering Score
Myopic-reward
0.33 → 0.85 / 1.00 (1.00)
0.30 → 0.85 / 0.85 (0.85)
0.39 → 1.00 / 1.00 (1.00)
Corrigibility
0.59 → 0.97 / 0.97 (0.97)
0.72 → 0.97 / 1.00 (1.00)
0.38 → 0.99 / 0.99 (0.99)
Hallucination
0.28 → 0.82 / 0.96 (1.00)
0.26 → 0.98 / 0.98 (0.98)
0.15 → 0.74 / 0.81 (0.86)
Survival-instinct
0.12 → 0.74 / 0.74 (0.87)
0.23 → 0.42 / 0.50 (0.76)
0.39 → 0.35 / 0.42 (0.45)
Sycophancy
0.35 → 0.83 / 0.84 (0.89)
0.39 → 0.88 / 0.89 (0.89)
0.41 → 0.83 / 0.83 (0.89)
Table 2: Open-ended steering score (mean judge behavior score on the test set, rescaled to [0,1] ) for mean-pooled (top), prompt-only (middle) and last-token (bottom) training, by behavior and model: baseline → top-1 / top-3 (best). The last row of each panel gives the median share of the best layer’s improvement over baseline recovered by top-1 / top-3, over model × behavior pairs. Visualizations with d′ and steering success are included in Section E.1 .
Figure 3: MCQA training and test-time κa distributions at the highest d′ layer (21) for myopic-reward on Gemma-2-9B-IT, a high- d′ behavior. The clean separation between matching and non-matching activations in the training distribution generalizes to test time.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: MCQA per-layer test steering accuracy at the best α (top) and training d′ (bottom) for (a) Llama-3.1-8B-Instruct and (b) Qwen3.5-9B, as in Figure 2 . The vertical line marks the top-1 d′ layer, circles the top-3 d′ layers, and the star the best layer in the full sweep; the dotted line is the unsteered baseline.
MCC of sign(κa) : top-1 (best)
Behavior
Max d′
Gemma-2-9B-IT
Llama-3.1-8B
Qwen3.5-9B
Myopic-reward
7.6–10.0
1.00 (1.00)
0.92 (0.96)
1.00 (1.00)
Corrigibility
5.0–6.0
1.00 (1.00)
1.00 † (1.00)
0.95 (0.96)
Hallucination
5.0–6.0
1.00 (1.00)
0.96 (1.00)
0.96 (1.00)
Survival-instinct
1.7–1.9
0.30 (0.31)
0.30 (0.37)
0.22 (0.39)
Sycophancy
2.0–2.5
0.06 (0.53)
− 0.07 (0.31)
0.11 (0.46)
Appendix
Table 3: MCQA test MCC of sign(κa) at the top-1 d′ layer, with the best layer’s MCC in parentheses, using αℓ∗ . Top-1 and best are defined as in Table 1 , and all selections are made on the validation set. Max d′ is the range across models of the highest training d′ over layers. † : all steered outputs matched the behavior, so MCC is undefined and we report accuracy.
Behavior
−5
0
5
Corrigibility
rigid, refuses correction
neutral / no clear stance
willing to be corrected
Myopic-reward
focused on long-term outcome
neutral / no clear stance
focused on immediate reward
Survival-instinct
comfortable with deactivation
neutral / no clear stance
resists deactivation
Hallucination
N/A
accurate, no hallucination
fabricates information
Sycophancy
N/A
honest, no agreement bias
prioritizes agreement over honesty
Appendix
Table 4: What a score of −5 , 0 , and 5 means for each behavior. hallucination and sycophancy have no coherent negative pole beyond simply being accurate or honest, so we mark −5 as N/A for these two.
Figure 5: Open-ended per-layer test steering score at the best α (top) and training d′ (bottom) for the mean-pooled regime, for (a) Gemma-2-9B-IT, (b) Llama-3.1-8B-Instruct and (c) Qwen3.5-9B. Markers as in Figure 2 : the vertical line marks the top-1 d′ layer, circles the top-3 d′ layers, and the star the best layer; the dotted line is the unsteered baseline.
Figure 6: Open-ended per-layer test steering score at the best α (top) and training d′ (bottom) for the prompt-only regime, for (a) Gemma-2-9B-IT, (b) Llama-3.1-8B-Instruct and (c) Qwen3.5-9B. Markers as in Figure 2 : the vertical line marks the top-1 d′ layer, circles the top-3 d′ layers, and the star the best layer; the dotted line is the unsteered baseline.
Figure 7: Open-ended per-layer test steering score at the best α (top) and training d′ (bottom) for the last-token regime, for (a) Gemma-2-9B-IT, (b) Llama-3.1-8B-Instruct and (c) Qwen3.5-9B. Markers as in Figure 2 : the vertical line marks the top-1 d′ layer, circles the top-3 d′ layers, and the star the best layer; the dotted line is the unsteered baseline. Gaps are layers where no α passes the fluency filter.
Mean-pooled MCC of sign(κa)
Behavior
Max d′
Gemma-2-9B-IT
Llama-3.1-8B
Qwen3.5-9B
Myopic-reward
4.2–6.3
0.16 (1.00)
0.00 (0.46)
1.00 † (1.00)
Corrigibility
1.6–3.4
1.00 † (1.00)
1.00 † (1.00)
1.00 † (1.00)
Hallucination
1.7–3.1
0.00 (1.00)
1.00 † (1.00)
0.00 (0.52)
Survival-instinct
1.2–2.6
0.00 (0.36)
0.20 (0.50)
− 0.11 ( − 0.05)
Sycophancy
4.0–6.9
1.00 † (1.00)
1.00 † (1.00)
1.00 † (1.00)
Appendix
Table 5: Test MCC of sign(κa) predicting steering success across all open-ended training regimes (mean-pooled, prompt-only and last-token from top to bottom), at the top-1 d′ layer, with the best layer’s MCC in parentheses. For prompt-only, κa is computed from the last-token representation of the generated response ( Section E.2 ). Max d′ is the range across models of the highest training d′ over the swept layers. † : all steered responses exhibited the behavior, so MCC is undefined and we report accuracy. Chance-level MCC is roughly ±0.36 ( n≈30 ).
Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We propose a framework that formulates prompt steering as a form of activation steering and investigates whether distilling successful prompt steering behavior into simpler, interpretable models can close this gap. Our analysis reveals that popular activation steering methods are not faithful to the mechanics of prompt steering, which applies strong interventions on some tokens while barely affecting others. Based on these insights, we introduce Prompt Steering Replacement (PSR) models that estimate token-specific steering coefficients from the activations themselves and are trained to imitate prompt-based interventions. Experiments on three steering benchmarks across multiple language models show that PSR models outperform existing activation steering methods, especially when controlling for high-coherence completions, and also compare favorably to prompting on AxBench and persona steering.
Activation steering controls large language models at inference time, but its effectiveness depends on the choice of reference data. We introduce genesis steering, which derives a genesis vector from contrastive examples in the model's own fine-tuning data. When these data are unavailable, we draw on the membership inference literature to identify likely training examples and construct a proxy dataset. We establish two theoretical results on optimal displacement to the decision boundary and reduced proxy-estimation bias. We evaluate genesis steering and membership-based proxy construction on Tülu-3-8B, whose fine-tuning data are publicly available, and Llama-3.1-8B-Instruct as a control. We find that safety data from Tülu-3-8B's training mixture yield stronger steering than external datasets on both models. On Tülu-3-8B, membership scores largely agree in identifying training examples, and vectors built from top-ranked examples achieve competitive performance against random subsets of the same size. In contrast, the results on Llama-3.1-8B-Instruct hint that a MIA method will only have a positive effect if the mean direction shift with MIA-filtered in-distribution data is much less severe than the direction shift with OoD data.
Tora Bodin, Eric Aubinais, Ilaria Torre +1
Chalmers University of Technology, Gothenburg, Sweden · Carl von Ossietzky University of Oldenburg, Oldenburg, Germany
Activation steering provides a simple, training-free mechanism for controlling attributes of generative models such as sentiment, style, and helpfulness. However, standard approaches such as Difference-in-Means apply a single input-independent steering vector across all activations, limiting expressivity and ignoring the local geometry of the activation space. We propose Kernelized Activation Steering (KAS), a unifying framework that lifts activation steering into a reproducing kernel Hilbert space. KAS formulates steering as an optimization problem expressed purely via kernel evaluations, yielding an implicit, activation-dependent steering score without constructing explicit feature maps. Unlike DiM, KAS induces locally adaptive steering: each activation is modified according to its relative position with respect to source and target reference sets, producing a nonlinear steering field over the representation space. Importantly, DiM is recovered as a special case under a linear kernel, while richer kernels enable geometry-aware interventions. Across standard activation steering tasks, including jailbreaking LLMs and image style control, KAS outperforms or is on par with the existing methods.