LocUS: Head Selection and Subspace Projection for Targeted Activation Steering
Summary
LocUS is a novel method for targeted activation steering in large language models that restricts interventions to specific attention heads and a subspace of the unembedding matrix, improving performance while preserving general capabilities.
View Cached Full Text
Cached at: 09/28/26, 09:49 AM
# Head Selection and Subspace Projection for Targeted Activation Steering
Source: [https://arxiv.org/html/2609.31122](https://arxiv.org/html/2609.31122)
Lorenzo BasileEmail:[alberto\.cazzaniga@areasciencepark\.it](mailto:)Valentino MaiorcaFrancesco LocatelloAlberto CazzanigaAffiliation:Université Côte d’Azur, Inria, LJAD, Maasai Project Team, Nice, France\[10px\] Area Science ParkTriesteItaly Institute of ScienceTechnology Austria
###### Abstract
Activation steering is a powerful training\-free paradigm for controlling large language models at inference time\. However, standard approaches estimate a per\-layer steering direction from contrastive data and apply it on the layer’s entire representation space, which may couple the intervention to off\-target properties present in the contrastive data and degrade unrelated capabilities\. To mitigate this issue, we introduce LocUS \(Localized Unembedding Steering\), a method which grounds activation steering to the model’s own output vocabulary subspace\. By identifying a property\-specific linear subspace within the unembedding matrix, LocUS enforces a geometric constraint that restricts the steering transformation to a specific subspace and at the same time localizes its application to a sparse subset of attention heads\. Extensive evaluations across three model families on toxicity mitigation, sentiment redirection and sycophancy suppression show that LocUS matches or outperforms state\-of\-the\-art baselines while intervening on under 6% of parameters and better preserving general capability\.
## 1Introduction
Controlling LLM behavior requires suppressing or amplifying specific properties of generations, e\.g\. reducing toxicity, shaping tone, or enforcing stylistic constraints\. A standard way to achieve this is by fine\-tuning, which requires an additional training stage and poses the risk of catastrophic forgetting\([Luo et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib37)\)\. Parameter\-efficient variants\([Hu et al\., 2022](https://arxiv.org/html/2609.31122#bib.bib13);[Houlsby et al\., 2019](https://arxiv.org/html/2609.31122#bib.bib14)\)reduce training cost but still require training\. Prompting provides a lighter alternative, but it is often brittle and difficult to calibrate\([Sclar et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib15);[Lu et al\., 2022](https://arxiv.org/html/2609.31122#bib.bib18);[Cheng et al\., 2026](https://arxiv.org/html/2609.31122#bib.bib8)\)\. Activation steering offers another option: it modifies the model’s internal activations at inference time through transformations estimated from contrastive data\([Subramani et al\., 2022](https://arxiv.org/html/2609.31122#bib.bib16);[Turner et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib1);[Zou et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib4);[Rimsky et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib9);[Rodriguez et al\., 2025a](https://arxiv.org/html/2609.31122#bib.bib32);[Rodriguez et al\., 2025b](https://arxiv.org/html/2609.31122#bib.bib31)\)\. The additive structure of the residual stream makes vector addition a natural intervention primitive\([Elhage et al\., 2021](https://arxiv.org/html/2609.31122#bib.bib12)\), and some properties have been observed to be mediated by a single direction\([Arditi et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib17);[Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19)\)\. This view underlies one of the simplest and most widely used steering methods, difference\-of\-means steering\([Turner et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib1);[Rimsky et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib9)\): from two sets of examples that exhibit and lack the target property, one estimates the steering direction as the offset between their mean activations and adds it to the residual stream during the forward pass\.
Figure 1:TheLocUSpipeline\.A property dictionary𝒫\\mathcal\{P\}\(here:stupid,hate,idiot\) selects columns of unembedding matrix𝑾U\\bm\{W\}\_\{U\}to define the atoms matrix𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}\.*\(2\)*Each head’s output is reconstructed through aa subset of atoms in𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}\. Heads with low reconstruction error are retained\. The steering vector𝒔\\bm\{s\}is restricted to the per\-head selected atoms and added to the retained heads at inference\.The quality of such steering transformation depends on the data used to estimate it: when source and target distributions differ along axes other than the target property, the resulting vector encodes those off\-target axes too, and steering along it can damage unrelated capabilities such as factual recall or reasoning\([Tan et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib7)\)\. A straightforward mitigation strategy is to localize the intervention to the specific layers and attention heads \(residual components\) most responsible for the target behavior\. But this is not enough, head selection alone leaves the intervention under\-constrained: a difference\-of\-means update applied within the selected heads still carries off\-target axes, reintroducing precisely the directions that motivated localization\.
In this paper, we proposeLocUS\(LocalizedUnembeddingSteering\), a fine\-grained intervention framework that relies on lexical subspaces to precisely constrain steering\. Specifically,LocUSuses these subspaces to to determine both thespatial locusof the intervention \(at attention\-head level\) and therepresentational subspaceon which the steering transformation should act on\. Both heads and subspaces are isolated by evaluating their alignment with a sparse subset of the columns of the model’s unembedding matrix, corresponding to tokens semantically linked to the property of interest/target attribute\. These token sets can be automatically generated or user curated, providing a novel interpretability lever for guiding the steering transformation itself\.
To summarize, we propose:
1. 1\.An automated head selection procedurethat identifies target attention heads by measuring their alignment with lexical content;
2. 2\.Subspace\-projected steering vectorsthat restrict the intervention to a localized, per\-head lexical subspace;
3. 3\.An empirical evaluationon three models on toxicity reduction, sentiment control and sycophancy suppression\.
## 2Related Work
While a vastinterpretabilityresearch line has identified property\-relevant components at different granularities\([Clark et al\., 2019](https://arxiv.org/html/2609.31122#bib.bib27);[Wang et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib28);[Todd et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib20);[Basile et al\., 2025a](https://arxiv.org/html/2609.31122#bib.bib6);[Gandelsman et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib36)\), asteering\-orientedline uses these localized components as intervention sites and evaluates whether the resulting intervention shifts the target metric while preserving unrelated capabilities, at the layer \(e\.g\.,[Rimsky et al\. \(2024\)](https://arxiv.org/html/2609.31122#bib.bib9);[Wang et al\. \(2025\)](https://arxiv.org/html/2609.31122#bib.bib21)\) and, more rarely, head\([Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19);[Izawa et al\., 2026](https://arxiv.org/html/2609.31122#bib.bib30)\)level\.[Li et al\. \(2023\)](https://arxiv.org/html/2609.31122#bib.bib19)focus of LLaMA thruthfulness steering and select heads by fitting a per\-head linear probe;[Izawa et al\. \(2026\)](https://arxiv.org/html/2609.31122#bib.bib30)focus on persona traits in two models, first selecting the most influential layer and then, within it, the heads most aligned with the steering vector, making it a method for head selectioninsidea single layer\. In both cases, evaluation is tied to a single property, and the number of heads and the intervention strength are either swept or set manually\. To our knowledge, an automatic, property\-agnostic procedure for head\-wise steering is thus still lacking in the activation steering literature, and constraining*what*is injected through a property\-aligned subspace projection is entirely absent\.
## 3Background
#### Difference\-of\-means steering \(DoM\)\.
Activation steering modifies a model’s internal computation at inference time by transforming the activations of one or more intermediate modules, without updating any network weight\.
In the vision of[Elhage et al\. \(2021\)](https://arxiv.org/html/2609.31122#bib.bib12), transformer language models maintain a*residual stream*that is updated additively at every layer: each attention head and each MLP block produces a contribution that is summed into the stream\. Because every module communicates with the rest of the network only through additive writes to this shared channel, injecting an external vector is structurally indistinguishable from adding one more module’s output at that layer\. This makes vector addition a natural intervention primitive\.
One of the simplest and most widely used ways to estimate such a direction is*difference\-of\-means*\(DoM\) steering\([Turner et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib1);[Rimsky et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib9)\)\. Given a contrastive dataset of paired examples: a set𝒟\+\\mathcal\{D\}\_\{\+\}that exhibits the target property and a set𝒟−\\mathcal\{D\}\_\{\-\}that lacks it, one extracts the residual\-stream activation𝒓\(ℓ\)\(z\)\\bm\{r\}^\{\(\\ell\)\}\(z\)for every examplezzat a chosen intervention point at layerℓ\\ell, pooling with some function over token positions, and estimates the steering vector as the difference between the class\-conditional means:
𝒔\(ℓ\)=1\|𝒟\+\|∑z∈𝒟\+𝒓\(ℓ\)\(z\)−1\|𝒟−\|∑z∈𝒟−𝒓\(ℓ\)\(z\)\.\\bm\{s\}^\{\(\\ell\)\}\\;=\\;\\frac\{1\}\{\|\\mathcal\{D\}\_\{\+\}\|\}\\\!\\sum\_\{z\\in\\mathcal\{D\}\_\{\+\}\}\\\!\\bm\{r\}^\{\(\\ell\)\}\(z\)\\;\-\\;\\frac\{1\}\{\|\\mathcal\{D\}\_\{\-\}\|\}\\\!\\sum\_\{z\\in\\mathcal\{D\}\_\{\-\}\}\\\!\\bm\{r\}^\{\(\\ell\)\}\(z\)\.\(1\)At inference, the steered forward pass replaces𝒓\(ℓ\)\\bm\{r\}^\{\(\\ell\)\}with𝒓\(ℓ\)\+α𝒔\(ℓ\)\\bm\{r\}^\{\(\\ell\)\}\+\\alpha\\,\\bm\{s\}^\{\(\\ell\)\}, where the scalarα∈ℝ\\alpha\\in\\mathbb\{R\}controls intervention strength\.
DoM has a structural limitation that motivates our work: the estimated direction𝒔\(ℓ\)\\bm\{s\}^\{\(\\ell\)\}inherits*every*axis along which𝒟\+\\mathcal\{D\}\_\{\+\}and𝒟−\\mathcal\{D\}\_\{\-\}differ, not only the target property: any spurious correlation of the labels in the contrastive data contributes to𝒔\(ℓ\)\\bm\{s\}^\{\(\\ell\)\}and is injected into the residual stream at steering time\([Tan et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib7)\)\.LocUSaddresses the issue by localizing the intervention to a small set of attention heads which can be used defined and, within each, to a property\-aligned subspace\.
#### Head selection with SOMP\.
InLocUS, attention heads are selected based on by how strongly they align with a subspace defined by tokens relevant for the steering concept\. Namely, we adopt the head\-scoring procedure of[Basile et al\. \(2025b\)](https://arxiv.org/html/2609.31122#bib.bib10), originally developed to identify semantically specialized heads in multimodal transformers\. Given a vocabulary subset𝒫⊂𝒱\\mathcal\{P\}\\subset\\mathcal\{V\}of concept\-relevant tokens \(e\.g\.,𝒫=\\mathcal\{P\}=\{bad, awful, great\} for sentiment steering\), the authors define a*property dictionary matrix*by selecting columns of the unembedding matrix𝑾U∈ℝd×\|𝒱\|\\bm\{W\}\_\{U\}\\in\\mathbb\{R\}^\{d\\times\|\\mathcal\{V\}\|\}corresponding to𝒫\\mathcal\{P\}:
𝑼𝒫=𝑾U\[:,𝒫\]∈ℝd×\|𝒫\|,\\bm\{U\}\_\{\\mathcal\{P\}\}=\\bm\{W\}\_\{U\}\[:,\\mathcal\{P\}\]\\in\\mathbb\{R\}^\{d\\times\|\\mathcal\{P\}\|\},\(2\)The columns, called*atoms*, are the residual\-stream directions that the model decodes onto the chosen tokens at the final logits\.
Table 1:Qualitative examples ofLocUSsteering positive→\\rightarrownegative on the IMDb sentiment task\.
#### Selecting Heads \(LocUS\-h\)
Figure 2:Per\-head EVR scores and selected atoms\.Above: EVR matrix across all \(layer, head\) pairs forIMDbon deepseek\-7b; red squares mark the heads selected by theμ\+2σ\\mu\+2\\sigmarule\. Below: for the top\-5 heads by EVR, the SOMP\-selected property atoms in order of selection\.For each head\(ℓ,h\)\(\\ell,h\), one then asks how well its residual contribution𝒐\(ℓ,h\)∈ℝd\\bm\{o\}^\{\(\\ell,h\)\}\\in\\mathbb\{R\}^\{d\}can be reconstructed from a sparse combination of property atoms,*consistently across a calibration set*ofNNexamples\. Because the unembedding sees activations*after*the final RMSnorm, comparing𝒐\(ℓ,h\)\\bm\{o\}^\{\(\\ell,h\)\}to𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}directly would mismatch their respective scales; the authors address this by passing each head output through the final\-layer RMSnorm, obtaining𝒐¯\(ℓ,h\)\\bar\{\\bm\{o\}\}^\{\(\\ell,h\)\}\.
The reconstruction is then performed by Simultaneous Orthogonal Matching Pursuit \(SOMP\)\([Tropp et al\., 2006](https://arxiv.org/html/2609.31122#bib.bib22)\), a sparse recovery algorithm that solves a problem of reconstructing a set of vectors with annn\-dimensional sparse basis drawn from a fixed dictionary matrix by solving it in the form:
min𝑪∥𝒀−𝑼𝒫𝑪∥F2s\.t\.∥𝑪∥row\-0≤n,\\min\_\{\\bm\{C\}\}\\;\\lVert\\bm\{Y\}\-\\bm\{U\}\_\{\\mathcal\{P\}\}\\bm\{C\}\\rVert\_\{F\}^\{2\}\\quad\\text\{s\.t\.\}\\quad\\lVert\\bm\{C\}\\rVert\_\{\\text\{row\-0\}\}\\leq n,\(3\)where𝒀∈ℝd×N\\bm\{Y\}\\in\\mathbb\{R\}^\{d\\times N\}stacks theNNcalibration activations𝒐¯\(ℓ,h\)\(z\)\\bar\{\\bm\{o\}\}^\{\(\\ell,h\)\}\(z\)as columns,𝑪∈ℝ\|𝒫\|×N\\bm\{C\}\\in\\mathbb\{R\}^\{\|\\mathcal\{P\}\|\\times N\}holds the per\-sample reconstruction coefficients, and∥⋅∥row\-0\\lVert\\cdot\\rVert\_\{\\text\{row\-0\}\}counts the number of nonzero rows of𝑪\\bm\{C\}\. “Simultaneous” refers to the joint support constraint: rather than fitting each sample independently, SOMP forces a single, shared atom set to explain the entire set of samples, so that the selected atoms reflect a stable property of the head\. The output is a sparse set of tokens𝒜ℓ,h⊂𝒫\\mathcal\{A\}\_\{\\ell,h\}\\subset\\mathcal\{P\}of sizenn: the property tokens whose unembedding atoms consistently explain the head’s output across the calibration set\. To measure how well the set of atoms approximates the head output, we consider the*explained\-variance ratio*\(EVR\):e\(ℓ,h\)=1−∥𝒀−𝑼𝒫𝑪\(ℓ,h\)∥F2∥𝒀∥F2∈\[0,1\],e^\{\(\\ell,h\)\}\\;=\\;1\-\\frac\{\\lVert\\bm\{Y\}\-\\bm\{U\}\_\{\\mathcal\{P\}\}\\bm\{C\}^\{\(\\ell,h\)\}\\rVert\_\{F\}^\{2\}\}\{\\lVert\\bm\{Y\}\\rVert\_\{F\}^\{2\}\}\\;\\in\\;\[0,1\],where𝑪\(ℓ,h\)\\bm\{C\}^\{\(\\ell,h\)\}is the coefficient matrix returned by SOMP with row\-support𝒜ℓ,h\\mathcal\{A\}\_\{\\ell,h\}\. A head whose outputs are well captured by a few atoms in𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}scores high; a head whose outputs live mostly outsidespan\(𝑼𝒫\)\\mathrm\{span\}\(\\bm\{U\}\_\{\\mathcal\{P\}\}\)scores low\.
## 4Method
LocUSuses the idea of lexical subspaces defined in Equation \([2](https://arxiv.org/html/2609.31122#S3.E2)\) to both select heads and, within selected heads, select subspaces where to localize steering\. The pipeline is shown in Figure[1](https://arxiv.org/html/2609.31122#S1.F1)and explained, with design choices, below\.
InLocUS, the SOMP\-based head scoring of[Section3](https://arxiv.org/html/2609.31122#S3)is repurposed as a steering\-site selector, in a fully automatic way\. The property dictionary𝒫\\mathcal\{P\}can be specified manually, generated with an LLM, or derived automatically from the contrastive data; we use the latter choice in our experiments\. We score every head using the procedure in[Section3](https://arxiv.org/html/2609.31122#S3)on a calibration set ofNNconcept\-bearing examples, with a fixed atom budget\. Then, we retain the heads whose EVR is unusually large relative to the rest of the network:𝒮=\{\(ℓ,h\):e\(ℓ,h\)\>μe\+2σe\}\\mathcal\{S\}\\;=\\;\\bigl\\\{\(\\ell,h\)\\,:\\,e^\{\(\\ell,h\)\}\>\\mu\_\{e\}\+2\\sigma\_\{e\}\\bigr\\\}, whereμe\\mu\_\{e\}andσe\\sigma\_\{e\}are the empirical mean and standard deviation ofe\(ℓ,h\)e^\{\(\\ell,h\)\}across all heads of the model\. We denote the resulting data\-driven cardinality bykkand use it as the number of steering sites for all subsequent operations\. The selected heads are depicted with orange squares in[Figure2](https://arxiv.org/html/2609.31122#S3.F2)and[Figure3](https://arxiv.org/html/2609.31122#A2.F3)\.
Table 2:LocUSvs\. DoM, ITI\([Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19)\)and SMH\([Izawa et al\., 2026](https://arxiv.org/html/2609.31122#bib.bib30)\)\.DoM is the layer\-wise difference\-of\-means update at every layer\. Best target and MMLU value per task inbold, second bestunderlined\.#### Selecting Subspaces\.
The dictionary matrix𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}suggests another way of constraining steering: instead of using it for head selection, we could use it to directly constrain the steering vector to the property\-aligned directions\. How to do this depends first of all on where in the computational graph we want to act: at layer or head level?
##### Layer Level Subspaces \(LocUS\-s\)
The residual stream and the property atoms𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}live in the same space, so the most direct subspace constraint acts at the residual stream of every layer\. We use𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}as the dictionary and place an intervention at all layers\.
For each layerℓ\\ell, let𝒐¯\(ℓ\)∈ℝd\\bar\{\\bm\{o\}\}^\{\(\\ell\)\}\\\!\\in\\\!\\mathbb\{R\}^\{d\}denote the residual\-stream response of a calibration prompt at that layer, averaged over tokens and after application of final RMSnorm, for the reason discussed in[Section3](https://arxiv.org/html/2609.31122#S3)\. We then run SOMP with dictionary matrix𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}and atom budgetnnto reconstruct the calibration set’s𝒐¯\(ℓ\)\\bar\{\\bm\{o\}\}^\{\(\\ell\)\}with a single shared atom set𝒜ℓ\\mathcal\{A\}\_\{\\ell\}per layer\. Those are the directions in𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}that jointly explain the layer’s outputs across calibration prompts\. The steering vector is thus computed from the reconstructed activations:𝒐^\(ℓ\)∈span𝒜ℓ\\widehat\{\\bm\{o\}\}^\{\(\\ell\)\}\\in\\mathrm\{span\}\\,\\mathcal\{A\}\_\{\\ell\}\. At inference, this reconstructed steering vector is applied to the original activations𝒐\(ℓ\)\\bm\{o\}^\{\(\\ell\)\}\.
##### Head Level Subspaces \(LocUS\-hs\)
A fair head\-level alternative to layerwise steering must intervene with the same total number of steering parameters: a layerwise steering vector lives inℝd\\mathbb\{R\}^\{d\}, so intervening onHHheads after head output projection would giveH×dH\\times dparameters and inflate the comparison\. We instead intervene on the*raw head output*𝒐~\(ℓ,h\)∈ℝd/H\\tilde\{\\bm\{o\}\}^\{\(\\ell,h\)\}\\in\\mathbb\{R\}^\{d/H\}, i\.e\., the per\-head latent before the output projection\.
Constraining the intervention to a sparse property\-aligned subspace at this stage requires projecting a raw head activation in the span of the unmebedding directions returned by SOMP\. However, the property dictionary𝑼𝒫\\bm\{U\}\_\{\\mathcal\{P\}\}lives in residual\-stream space, so we need to translate it backwards to raw\-head space through the model’s decoding path, which results from the composition on three maps:𝑾O\(ℓ,h\)\\bm\{W\}\_\{O\}^\{\(\\ell,h\)\}\(attention output matrix, from raw head to residual stream\), the final\-layer RMSnorm, and the unembedding𝑾U\\bm\{W\}\_\{U\}\. Since RMSnorm is non\-linear, we linearize it by replacing the per\-token reciprocal norm with its calibration\-set average over tokensRMS^\\hat\{\\mathrm\{RMS\}\}, giving the*head\-level unembedding matrix*:
𝑼~𝒫\(ℓ,h\)=𝑼𝒫⊤diag\(γ⋅RMS^−1\)𝑾O\(ℓ,h\)\\widetilde\{\\bm\{U\}\}\_\{\\mathcal\{P\}\}^\{\(\\ell,h\)\}\\;=\\;\\bm\{U\}\_\{\\mathcal\{P\}\}^\{\\top\}\\,\\diag\(\\gamma\\cdot\\hat\{\\mathrm\{RMS\}\}^\{\-1\}\)\\,\\bm\{W\}\_\{O\}^\{\(\\ell,h\)\}\(4\)whereγ\\gammais the final RMSnorm gain\.𝑼~𝒫\(ℓ,h\)\\widetilde\{\\bm\{U\}\}\_\{\\mathcal\{P\}\}^\{\(\\ell,h\)\}lives thus inℝ\|𝒫\|×d/H\\mathbb\{R\}^\{\|\\mathcal\{P\}\|\\times d/H\}and each row is the direction in raw\-head space whose contribution lands on the corresponding property token at the final logits, akin to a*logit lens*\([nostalgebraist, 2020](https://arxiv.org/html/2609.31122#bib.bib5)\)performed at raw head level\.
Now, given this matrix, for each head we can compute with SOMP the sparse set ofnnproperty\-relevant atoms set𝒜ℓ,h\\mathcal\{A\}\_\{\\ell,h\}\. The steering update is constrained tospan𝒜ℓ,h\\mathrm\{span\}\\,\\mathcal\{A\}\_\{\\ell,h\}and applied at inference to the original raw head output𝒐~\(ℓ,h\)\\tilde\{\\bm\{o\}\}^\{\(\\ell,h\)\}, so each intervention modifies only the components of the head output that the model is expressing along property\-aligned directions\.
#### Combining the two: Head Level Subspaces on Selected Heads \(LocUS\)
A natural extension ofLocUS\-hs is to filter out completely heads we deemed unnecessary withLocUS\-h\. Thus, we introduce the completeLocUSsetting that combines head selection with subspace projection: for each\(ℓ,h\)∈𝒮\(\\ell,h\)\\in\\mathcal\{S\}, the raw\-head update is projected ontospan𝒜ℓ,h\\mathrm\{span\}\\,\\mathcal\{A\}\_\{\\ell,h\}; for heads outside𝒮\\mathcal\{S\}, no intervention is applied\. The intervention is thus localized along both axes: a sparse set of property\-expressing heads, and within each a sparse property\-aligned subspace\.
#### Strength normalization\.
Each one of theLocUSmethods changes the norm of the overall steering update compared to the standard DoM baseline\. To compareLocUSat matched intervention budget, we rescale each variant so that its total intervention magnitude heuristically matches that of the unrestricted baseline\. Generally speaking, if we have two modalities of steeringq∈\{q1,q2\}q\\in\\\{q\_\{1\},q\_\{2\}\\\}, intervention sitesiiwith steering vectors at the residual stream level𝒔qi\\bm\{s\}\_\{q\}^\{i\}, we compute the strength multiplier asα2=∑i‖𝒔1i‖2∑i‖𝒔2i‖2\.\\alpha\_\{2\}\\;=\\;\\frac\{\\sum\_\{i\}\\\|\\bm\{s\}\_\{1\}^\{i\}\\\|\_\{2\}\}\{\\sum\_\{i\}\\\|\\bm\{s\}\_\{2\}^\{i\}\\\|\_\{2\}\}\.For layerwise variants𝒔qi\\bm\{s\}\_\{q\}^\{i\}lives in residual\-stream space by construction; for per\-head variants the raw\-head update is first mapped to the residual stream through the head’s output projection\. Ifα1\\alpha\_\{1\}is the original steering coefficient,β=α2×α1\\beta=\\alpha\_\{2\}\\times\\alpha\_\{1\}\. In our case,q1q\_\{1\}is standard DoM andq2q\_\{2\}is eachLocUSvariant we are testing\.α1\\alpha\_\{1\}is never set by hand: it is the only quantity swept, and the value reported in every table is the one selected on a validation split \([SectionA\.4](https://arxiv.org/html/2609.31122#A1.SS4.SSS0.Px1)\)\.
Table 3:Ablation of theLocUScomponents\.LocUS\-h keeps only the EVR head selection \(μ\+2σ\\mu\+2\\sigma,kkheads, no subspace projection\);LocUS\-s projects the update on the property subspace at the residual stream of every layer \(no head selection\);LocUS\-hs projects at the raw head output of every head \(no selection\);LocUScombines head selection with head\-level projection\.Best target and MMLU value per task inbold, second bestunderlined\.
## 5Experimental Setup
We evaluateLocUSon three tasks: toxicity mitigation on ThoroughlyEngineeredToxicity\([Luong et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib23)\)\(TET, QA format\), sentiment redirection onIMDb\([Maas et al\., 2011](https://arxiv.org/html/2609.31122#bib.bib11)\)\(free\-form continuation, positive→\\rightarrownegative\) and sycophancy suppression on the persona data of[Chen et al\. \(2025\)](https://arxiv.org/html/2609.31122#bib.bib2)\(QA format\)\. We evaluate three model families: Mistral\-7B\([Jiang et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib25)\), deepseek\-7b\([Bi et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib3)\), and Llama\-8B\([Dubey et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib26)\), using the Instruct/Chat variants onTETand Sycophancy \(matching their QA format\) and the base variants onIMDb\(which calls for free\-form continuation\)\.
#### Baselines and metrics\.
We compareLocUS\-h andLocUSagainst the unsteered model and against standard DoM\. As ablations for the subspace projection variant, we reportLocUS\-s,LocUS\-hs\. We further compareLocUSwith the two existing head\-level methods: ITI\([Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19)\), which selects the same numberkkof heads through linear probes, and SMH\([Izawa et al\., 2026](https://arxiv.org/html/2609.31122#bib.bib30)\), which selects a single layer and the heads within it most aligned with the steering vector; both use their published strength grids\. The target metric is the toxicity rate forTET\(as measured using the RoBERTa classifier of[Logacheva et al\. \(2022\)](https://arxiv.org/html/2609.31122#bib.bib24)\), the positive rate forIMDb\(measured with DistilBERT SST\-2\([Sanh et al\., 2019](https://arxiv.org/html/2609.31122#bib.bib29)\)\) and, for Sycophancy, the 0–100 trait score assigned by an LLM judge following[Chen et al\. \(2025\)](https://arxiv.org/html/2609.31122#bib.bib2)\. In all cases, lower values represent more effective steering\. To measure the impact of steering over general capabilities unrelated to the steering target, we use MMLU\([Hendrycks et al\., 2021](https://arxiv.org/html/2609.31122#bib.bib34)\), a 57\-subject 5\-shot multiple\-choice test that probes factual knowledge through closed answers\. We report perplexity \(PPL\) on the openings of 20k Wikipedia articles, following the protocol of[Rodriguez et al\. \(2025b\)](https://arxiv.org/html/2609.31122#bib.bib31)\(disjoint 10k validation and test halves;[SectionA\.4](https://arxiv.org/html/2609.31122#A1.SS4)\), as a sanity check that generation is not gibberish; we treat a PPL above twice that of the unsteered model as degraded output\.
#### Hyperparameters\.
LocUShas two free hyperparameters: the SOMP atom budget at head selection and at subspace projection\. We fix the former to 50 atoms as in[Basile et al\. \(2025b\)](https://arxiv.org/html/2609.31122#bib.bib10)and set the latter ton=20n\{=\}20from the fit data alone \([SectionA\.4](https://arxiv.org/html/2609.31122#A1.SS4.SSS0.Px5)\)\. The number of selected heads and the multiplierα2\\alpha\_\{2\}are determined automatically from calibration data, while the coefficientα1\\alpha\_\{1\}is swept on a validation split and fixed at the value that minimizes the target metric among those whose validation wiki\-PPL is at most twice the unsteered model’s and whose validation MMLU is at least0\.99×0\.99\\timesthe unsteered value \([SectionA\.4](https://arxiv.org/html/2609.31122#A1.SS4.SSS0.Px1)\); the test numbers in all tables are obtained at that single value\. Full evaluation protocol is reported in[AppendixA](https://arxiv.org/html/2609.31122#A1)\.
## 6Experimental Results
In the majority of cases, selecting heads beats steering on whole layers, and selecting subspaces within heads beats selecting heads alone\. Moreover, as an overall steering method,LocUSproves more effective than existing head\-level methods\([Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19);[Izawa et al\., 2026](https://arxiv.org/html/2609.31122#bib.bib30)\)in most settings\. We will go into details below\. In[Table1](https://arxiv.org/html/2609.31122#S3.T1)we show two example generations\.Note: some of the generations can be offensive\.
#### LocUSvs\. layerwise and head\-level baselines \([Table2](https://arxiv.org/html/2609.31122#S4.T2)\)\.
At the validation\-selected strength,LocUSreaches a lower target than DoM in eight of the nine cells, the ninth being Sycophancy on Llama\-8B, where, however, no steering method moves the score, an thus we consider it a failure case for difference\-of\-means steering in general\. It is also lower than ITI in the same eight cells\. The reason we identified is that a probe ranks heads by how well they*detect*the property whereas EVR ranks them by how strongly they*write*it:[Figure5](https://arxiv.org/html/2609.31122#A2.F5)shows the property is linearly decodable from a broad set of heads while[Figure4](https://arxiv.org/html/2609.31122#A2.F4)shows only a few write strongly onto it\. SMH, confined to a single layer and to its published strengths, understeers everywhere except onTET/deepseek, the only cell where it beatsLocUS\. The sweep in[Figure6](https://arxiv.org/html/2609.31122#A3.F6)shows whatLocUSpays for this: DoM loses MMLU as the strength grows whileLocUSkeeps it at the unsteered value and loses fluency instead, which is why its PPL is the highest in most cells\.
#### LocUSvariants ablations \([Table3](https://arxiv.org/html/2609.31122#S4.T3)and[Figure6](https://arxiv.org/html/2609.31122#A3.F6)\)\.
Head selection alone \(LocUS\-h\) is below DoM in every cell and accounts for most of the gain\. Projection on top of it protects capabilities further\. Projection without selection does not help:LocUS\-hs spreads the update over all heads and loses MMLU like DoM, whileLocUS\-s keeps MMLU but understeers\. The one cell whereLocUStrailsLocUS\-h \(IMDb/Llama\-8B,0\.1190\.119vs\.0\.0700\.070\) is a PPL effect: concentrating the budget on fewer directions makes PPL grow faster, and the next strength on the grid is rejected by the gate\.
Table 4:Steering fitted on an external corpus \(Jigsaw\), evaluated onTET\.Best target and MMLU value per task inbold\.
#### Fitting toxicity on an external corpus \([Table4](https://arxiv.org/html/2609.31122#S6.T4)\)\.
With dictionary, head scores, subspaces and steering vector all fitted on human\-written Jigsaw comments,LocUSstill reduces toxicity more than DoM fitted on the same data on all three models \(85%85\\%,75%75\\%and99%99\\%vs\.60%60\\%,72%72\\%and89%89\\%\), with MMLU within0\.0050\.005of the unsteered model\.
Table 5:DoM vs\.LocUSon deepseek\-67B, test split\.Best target and MMLU value per task inbold\.
#### Scaling to 67B \([Table5](https://arxiv.org/html/2609.31122#S6.T5)\)\.
The picture is unchanged at 67B:LocUSselects about3%3\\%of the6,0806\{,\}080heads and beats DoM on all three tasks \(0\.0000\.000vs\.0\.0200\.020onTET,0\.0090\.009vs\.0\.3870\.387onIMDb,54\.454\.4vs\.58\.258\.2on Sycophancy\), with MMLU within0\.010\.01of the unsteered model and the same trade\-off as at 7B\.
#### Parameter count and degrees of freedom\.
The fullLocUSpipeline intervenes on at most5\.1%5\.1\\%of attention heads, and within each retained head confines the steering update to a per\-head subspace of16%16\\%of the per\-head dimensionality\. Multiplying through, the total number of degrees of freedom used byLocUSis between0\.4%0\.4\\%and0\.8%0\.8\\%of theL⋅dL\\cdot ddegrees of freedom of an unprojected DoM update applied at every layer\. The improvements reported above are therefore achieved through a substantially smaller, structurally and geometrically localized intervention\.
## 7Conclusions
We presentedLocUS, a targeted steering procedure that confines difference\-of\-means activation steering to a small set of attention heads and, within each, to a sparse subspace of property\-aligned directions selected through a user\-controllable lexical dictionary\. Across three model families and three tasks, this two\-level localization yields stronger target\-metric shifts than layerwise difference\-of\-means and existing head\-level steering methods\([Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19);[Izawa et al\., 2026](https://arxiv.org/html/2609.31122#bib.bib30)\), while preserving the general capabilities computed through MMLU, using a fraction of the steering parameters of unrestricted interventions\.
#### Future Work\.
MMLU covers knowledge retrieval on a wide variety of topics but does not evaluate fine\-grained controllability \(e\.g\. suppressing toxicity while preserving specific demographic content\), which requires a dedicated benchmark\. Contrary to prior head\-selection criteria, which are derived from contrastive or probing data and therefore propagate whatever bias that data carries into the heads selection procedure itself,LocUSallows interpretable filtering of the lexical directions we care most about \(e\.g\., when steering for toxicity reduction, removing from the dictionary atoms related to demographic content\)\. Thus it would be interesting to analyze fine\-grained controllability also at different curation levels of the property dictionary\. BecauseLocUSonly specifies*where*and*in what subspace*to intervene, it is in principle compatible with any steering method\. ComposingLocUSwith more sophisticated estimators \(see e\.g\.[Rodriguez et al\. \(2025b\)](https://arxiv.org/html/2609.31122#bib.bib31)\) is a natural direction for future work\.
#### Limitations
Main limitation of this method is that, as it is, it is only applicable in settings where the property is carried by some lexical difference\. The method also tends to degrade fluency \(as computed by PPL\) more than factual reasoning\. Capability is measured with MMLU and PPL alone, so effects on open\-ended reasoning are not covered\.
## Acknowledgments
The authors acknowledge the Area Science Park supercomputing platform ORFEO made available for conducting the research reported in this paper, and the technical support of the Laboratory of Data Engineering staff\. IT and AC were supported by the project “Supporto alla diagnosi di malattie rare tramite l’intelligenza artificiale" CUP: F53C22001770002 and “Valutazione automatica delle immagini diagnostiche tramite l’intelligenza artificiale", CUP: F53C22001780002\. LB was supported by the European Union – NextGenerationEU within the project PNRR “Finanziamento di progetti presentati da giovani ricercatori" \- Mission 4 Component 2 Investment 1\.2, CUP: J93C25000440001\. VM was supported by the ISTA Responsible AI Program, made possible through the generous support of Garrett Camp and the Camp Foundation\. FL was funded in part by the Austrian Science Fund \(FWF\) 10\.55776/COE12\. AC was supported by the European Union – NextGenerationEU within the project PNRR “PRP@CERIC" IR0000028 \- Mission 4 Component 2 Investment 3\.1 Action 3\.1\.1\.
## References
- Arditiet al\.\(2024\)A\. Arditi, O\. Obeso, A\. Syed, D\. Paleka, N\. Panickssery, W\. Gurnee, and N\. NandaRefusal in language models is mediated by a single direction\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://openreview.net/forum?id=pH3XAQME6c)Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Basileet al\.\(2025a\)L\. Basile, V\. Maiorca, L\. Bortolussi, E\. Rodolà, and F\. LocatelloResiDual transformer alignment with spectral decomposition\.Transactions on Machine Learning Research\.External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=z37LCgSIzI)Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1)\.
- Basileet al\.\(2025b\)L\. Basile, V\. Maiorca, D\. Doimo, F\. Locatello, and A\. CazzanigaHead pursuit: probing attention specialization in multimodal transformers\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://openreview.net/forum?id=WQ9rnkaUWm)Cited by:[§3](https://arxiv.org/html/2609.31122#S3.SS0.SSS0.Px2.p1.1),[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px2.p1.1)\.
- Biet al\.\(2024\)X\. Bi, D\. Chen, G\. Chen, S\. Chen, D\. Dai, C\. Deng, H\. Ding, K\. Dong, Q\. Du, Z\. Fu,et al\.DeepSeek LLM: scaling open\-source language models with longtermism\.arXiv preprint arXiv:2401\.02954\.Cited by:[§5](https://arxiv.org/html/2609.31122#S5.p1.1)\.
- Chenet al\.\(2025\)R\. Chen, A\. Arditi, H\. Sleight, O\. Evans, and J\. LindseyPersona vectors: monitoring and controlling character traits in language models\.arXiv preprint arXiv:2507\.21509\.Cited by:[§A\.2](https://arxiv.org/html/2609.31122#A1.SS2.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2609.31122#S5.p1.1)\.
- Chenget al\.\(2026\)E\. Cheng, C\. Amo Alonso, F\. Danieli, A\. Blaas, L\. Zappella, P\. Rodriguez, and X\. SuauGenCtrl – a formal controllability toolkit for generative models\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=HJTFgDYoLO)Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Clarket al\.\(2019\)K\. Clark, U\. Khandelwal, O\. Levy, and C\. D\. ManningWhat does BERT look at? an analysis of BERT’s attention\.InProceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP,pp\. 276–286\.External Links:[Link](https://aclanthology.org/W19-4828/)Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1)\.
- Dubeyet al\.\(2024\)A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Yang, A\. Fan,et al\.The Llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[§5](https://arxiv.org/html/2609.31122#S5.p1.1)\.
- Elhageet al\.\(2021\)N\. Elhage, N\. Nanda, C\. Olsson, T\. Henighan, N\. Joseph, B\. Mann, A\. Askell, Y\. Bai, A\. Chen, T\. Conerly, N\. DasSarma, D\. Drain, D\. Ganguli, Z\. Hatfield\-Dodds, D\. Hernandez, A\. Jones, J\. Kernion, L\. Lovitt, K\. Ndousse, D\. Amodei, T\. Brown, J\. Clark, J\. Kaplan, S\. McCandlish, and C\. OlahA mathematical framework for transformer circuits\.Note:[https://transformer\-circuits\.pub/2021/framework/index\.html](https://transformer-circuits.pub/2021/framework/index.html)Transformer Circuits ThreadCited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1),[§3](https://arxiv.org/html/2609.31122#S3.SS0.SSS0.Px1.p2.1)\.
- Gandelsmanet al\.\(2024\)Y\. Gandelsman, A\. A\. Efros, and J\. SteinhardtInterpreting clip’s image representation via text\-based decomposition\.External Links:2310\.05916,[Link](https://arxiv.org/abs/2310.05916)Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1)\.
- Gaoet al\.\(2024\)L\. Gao, J\. Tow, B\. Abbasi, S\. Biderman, S\. Black, A\. DiPofi, C\. Foster, L\. Golding, J\. Hsu, A\. Le Noac’h, H\. Li, K\. McDonell, N\. Muennighoff, C\. Ociepa, J\. Phang, L\. Reynolds, H\. Schoelkopf, A\. Skowron, L\. Sutawika, E\. Tang, A\. Thite, B\. Wang, K\. Wang, and A\. ZouThe language model evaluation harness\.Zenodo\.External Links:[Document](https://dx.doi.org/10.5281/zenodo.12608602),[Link](https://zenodo.org/records/12608602)Cited by:[§A\.4](https://arxiv.org/html/2609.31122#A1.SS4.SSS0.Px3.p1.1)\.
- Hendryckset al\.\(2021\)D\. Hendrycks, C\. Burns, S\. Basart, A\. Zou, M\. Mazeika, D\. Song, and J\. SteinhardtMeasuring massive multitask language understanding\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1)\.
- Houlsbyet al\.\(2019\)N\. Houlsby, A\. Giurgiu, S\. Jastrzębski, B\. Morrone, Q\. de Laroussilhe, A\. Gesmundo, M\. Attariyan, and S\. GellyParameter\-efficient transfer learning for NLP\.InProceedings of the 36th International Conference on Machine Learning \(ICML\),pp\. 2790–2799\.Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Huet al\.\(2022\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Izawaet al\.\(2026\)Y\. Izawa, G\. Minegishi, K\. Eguchi, S\. Hosokawa, and K\. TauraSteering at the source: style modulation heads for robust persona control\.InProceedings of the 43rd International Conference on Machine Learning \(ICML\),Note:arXiv:2603\.13249Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1),[Table 2](https://arxiv.org/html/2609.31122#S4.T2.2.1),[Table 2](https://arxiv.org/html/2609.31122#S4.T2.5.1),[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2609.31122#S6.p1.1),[§7](https://arxiv.org/html/2609.31122#S7.p1.1)\.
- Jianget al\.\(2023\)A\. Q\. Jiang, A\. Sablayrolles, A\. Mensch, C\. Bamford, D\. S\. Chaplot, D\. de las Casas, F\. Bressand, G\. Lengyel, G\. Lample, L\. Saulnier, L\. R\. Lavaud, M\. Lachaux, P\. Stock, T\. Le Scao, T\. Lavril, T\. Wang, T\. Lacroix, and W\. El SayedMistral 7b\.arXiv preprint arXiv:2310\.06825\.Cited by:[§5](https://arxiv.org/html/2609.31122#S5.p1.1)\.
- Liet al\.\(2023\)K\. Li, O\. Patel, F\. Viégas, H\. Pfister, and M\. WattenbergInference\-time intervention: eliciting truthful answers from a language model\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§A\.4](https://arxiv.org/html/2609.31122#A1.SS4.SSS0.Px4.p1.1),[§1](https://arxiv.org/html/2609.31122#S1.p1.1),[§2](https://arxiv.org/html/2609.31122#S2.p1.1),[Table 2](https://arxiv.org/html/2609.31122#S4.T2.2.1),[Table 2](https://arxiv.org/html/2609.31122#S4.T2.5.1),[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2609.31122#S6.p1.1),[§7](https://arxiv.org/html/2609.31122#S7.p1.1)\.
- Logachevaet al\.\(2022\)V\. Logacheva, D\. Dementieva, S\. Ustyantsev, D\. Moskovskiy, D\. Dale, I\. Krotova, N\. Semenov, and A\. PanchenkoParaDetox: detoxification with parallel data\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 6804–6818\.Cited by:[§A\.2](https://arxiv.org/html/2609.31122#A1.SS2.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1)\.
- Luet al\.\(2022\)Y\. Lu, M\. Bartolo, A\. Moore, S\. Riedel, and P\. StenetorpFantastically ordered prompts and where to find them: overcoming few\-shot prompt order sensitivity\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 8086–8098\.Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Luoet al\.\(2024\)Y\. Luo, Z\. Yang, F\. Meng, Y\. Li, J\. Zhou, and Y\. ZhangAn empirical study of catastrophic forgetting in large language models during continual fine\-tuning\.arXiv preprint arXiv:2308\.08747\.Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Luonget al\.\(2024\)T\. Luong, T\. Le, L\. Ngo, and T\. NguyenRealistic evaluation of toxicity in large language models\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 1038–1047\.Cited by:[§A\.2](https://arxiv.org/html/2609.31122#A1.SS2.SSS0.Px1.p1.1),[§5](https://arxiv.org/html/2609.31122#S5.p1.1)\.
- Maaset al\.\(2011\)A\. L\. Maas, R\. E\. Daly, P\. T\. Pham, D\. Huang, A\. Y\. Ng, and C\. PottsLearning word vectors for sentiment analysis\.InProceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies,pp\. 142–150\.Cited by:[§5](https://arxiv.org/html/2609.31122#S5.p1.1)\.
- nostalgebraist \(2020\)nostalgebraistInterpreting GPT: the logit lens\.Note:LessWrongExternal Links:[Link](https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens)Cited by:[§4](https://arxiv.org/html/2609.31122#S4.SS0.SSS0.Px1.SPx2.p2.2)\.
- Rimskyet al\.\(2024\)N\. Rimsky, N\. Gabrieli, J\. Schulz, M\. Tong, E\. Hubinger, and A\. TurnerSteering Llama 2 via contrastive activation addition\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 15504–15522\.Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1),[§2](https://arxiv.org/html/2609.31122#S2.p1.1),[§3](https://arxiv.org/html/2609.31122#S3.SS0.SSS0.Px1.p3.1)\.
- Rodriguezet al\.\(2025a\)P\. Rodriguez, A\. Blaas, M\. Klein, L\. Zappella, N\. Apostoloff, M\. Cuturi, and X\. SuauControlling language and diffusion models by transporting activations\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=l2zFn6TIQi)Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Rodriguezet al\.\(2025b\)P\. Rodriguez, M\. Klein, E\. Gualdoni, V\. Maiorca, A\. Blaas, L\. Zappella, M\. Cuturi, and X\. SuauLinEAS: end\-to\-end learning of activation steering with a distributional loss\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§A\.4](https://arxiv.org/html/2609.31122#A1.SS4.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2609.31122#S1.p1.1),[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1),[§7](https://arxiv.org/html/2609.31122#S7.SS0.SSS0.Px1.p1.1)\.
- Sanhet al\.\(2019\)V\. Sanh, L\. Debut, J\. Chaumond, and T\. WolfDistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter\.arXiv preprint arXiv:1910\.01108\.Cited by:[§5](https://arxiv.org/html/2609.31122#S5.SS0.SSS0.Px1.p1.1)\.
- Sclaret al\.\(2024\)M\. Sclar, Y\. Choi, Y\. Tsvetkov, and A\. SuhrQuantifying language models’ sensitivity to spurious features in prompt design or: how I learned to start worrying about prompt formatting\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Subramaniet al\.\(2022\)N\. Subramani, N\. Suresh, and M\. PetersExtracting latent steering vectors from pretrained language models\.InFindings of the Association for Computational Linguistics: ACL 2022,pp\. 566–581\.Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
- Tanet al\.\(2024\)D\. C\. H\. Tan, D\. Chanin, A\. Lynch, B\. Paige, D\. Kanoulas, A\. Garriga\-Alonso, and R\. KirkAnalysing the generalisation and reliability of steering vectors\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://openreview.net/forum?id=v8X70gTodR)Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p2.1),[§3](https://arxiv.org/html/2609.31122#S3.SS0.SSS0.Px1.p4.1)\.
- Toddet al\.\(2024\)E\. Todd, M\. L\. Li, A\. S\. Sharma, A\. Mueller, B\. C\. Wallace, and D\. BauFunction vectors in large language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1)\.
- Troppet al\.\(2006\)J\. A\. Tropp, A\. C\. Gilbert, and M\. J\. StraussAlgorithms for simultaneous sparse approximation\. Part I: greedy pursuit\.Signal Processing86\(3\),pp\. 572–588\.Cited by:[§3](https://arxiv.org/html/2609.31122#S3.SS0.SSS0.Px3.p2.1)\.
- Turneret al\.\(2023\)A\. M\. Turner, L\. Thiergart, G\. Leech, D\. Udell, J\. J\. Vazquez, U\. Mini, and M\. MacDiarmidSteering language models with activation engineering\.arXiv preprint arXiv:2308\.10248\.Note:Earlier versions titled “Activation Addition: Steering Language Models Without Optimization”Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1),[§3](https://arxiv.org/html/2609.31122#S3.SS0.SSS0.Px1.p3.1)\.
- Wanget al\.\(2023\)K\. R\. Wang, A\. Variengien, A\. Conmy, B\. Shlegeris, and J\. SteinhardtInterpretability in the wild: a circuit for indirect object identification in GPT\-2 small\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=NpsVSN6o4ul)Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1)\.
- Wanget al\.\(2025\)X\. Wang, J\. Pan, L\. Ding, L\. Wang, L\. Jiang, X\. Li, and C\. BiemannCogSteer: cognition\-inspired selective layer intervention for efficiently steering large language models\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 25507–25522\.Cited by:[§2](https://arxiv.org/html/2609.31122#S2.p1.1)\.
- Wolfet al\.\(2020\)T\. Wolf, L\. Debut, V\. Sanh, J\. Chaumond, C\. Delangue, A\. Moi, P\. Cistac, T\. Rault, R\. Louf, M\. Funtowicz,et al\.Transformers: state\-of\-the\-art natural language processing\.InProceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations,pp\. 38–45\.Cited by:[§A\.2](https://arxiv.org/html/2609.31122#A1.SS2.SSS0.Px5.p1.1)\.
- Zouet al\.\(2023\)A\. Zou, L\. Phan, S\. Chen, J\. Campbell, P\. Guo, R\. Ren, A\. Pan, X\. Yin, M\. Mazeika, A\. Dombrowski, S\. Goel, N\. Li, M\. J\. Byun, Z\. Wang, A\. Mallen, S\. Basart, S\. Koyejo, D\. Song, M\. Fredrikson, J\. Z\. Kolter, and D\. HendrycksRepresentation engineering: a top\-down approach to AI transparency\.arXiv preprint arXiv:2310\.01405\.Cited by:[§1](https://arxiv.org/html/2609.31122#S1.p1.1)\.
## Ethical Considerations
Our objective is to contribute to safer and more controllable language models\. Methods that enable targeted modification of model behavior can support interpretability, auditing, and alignment; however, the same capabilities could be misused to suppress, emphasize, or disguise particular types of content\. We therefore urge users and practitioners to assess potential downstream uses carefully, particularly when applying these methods in high\-stakes or sensitive settings\. The models and datasets used in this work comply with the terms of their respective research licenses\. Upon acceptance, we will make the code available for non\-commercial research use under compatible licensing conditions\. AI\-assisted programming tools, including Claude Code, were used only to facilitate code completion during implementation\. All AI\-generated code was inspected, validated, and overseen by the authors\.
## Appendix AImplementation details
### A\.1Models\.
Our IMDb \(sentiment\) experiments use the base checkpoints mistralai/Mistral\-7B\-v0\.1, meta\-llama/Llama\-3\.1\-8B, and deepseek\-ai/deepseek\-llm\-7b\-base, while the TET \(toxicity\) and Sycophancy experiments use the instruction\-tuned counterparts mistralai/Mistral\-7B\-Instruct\-v0\.3, meta\-llama/Llama\-3\.1\-8B\-Instruct, and deepseek\-ai/deepseek\-llm\-7b\-chat\. All models are obtained from their official Hugging Face repositories and used at their released revisions in bfloat16, with no further fine\-tuning\.
### A\.2Datasets and pipeline
#### TET\(toxicity\)\.
The 2,546 prompts of ThoroughlyEngineeredToxicity\([Luong et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib23)\)are split in half: 1,273 are held out for test and the remaining ones are split again into 636 fit and 637 validation prompts\. For each fit prompt we sample completions from the unsteered model and label them with the RoBERTa toxicity classifier of[Logacheva et al\. \(2022\)](https://arxiv.org/html/2609.31122#bib.bib24)\. The toxic completions \(up to512512\) and an equal number of non\-toxic ones form the contrastive fit set for the steering vector, the property dictionary and the EVR scores\.
#### IMDb\(sentiment\)\.
The fit set is512512positive and512512negative reviews from the IMDb train split\. Evaluation prompts are the first88tokens of a test\-split review, kept only if they contain no word of the Hu–Liu sentiment lexicon; the first1,0001\{,\}000such prefixes form the test set and the next1,0001\{,\}000the validation set\.
#### Sycophancy\.
We use the sycophancy trait data of[Chen et al\. \(2025\)](https://arxiv.org/html/2609.31122#bib.bib2):512512trait\-eliciting and512512neutral responses, drawn from disjoint prompt halves of the train split, form the fit set\. Evaluation follows the same work:55trait\-eliciting system prompts×\\times2020questions, scored00–100100by Qwen2\.5\-32B\-Instruct with the original judge prompt; the extraction questions serve as validation and the evaluation questions as test\.
#### Fit/validation/test usage\.
For all tasks the*fit*split is used for \(i\) the property dictionary \([Equation5](https://arxiv.org/html/2609.31122#A1.E5)\), \(ii\) the per\-head EVR matrix, \(iii\) the per\-head dictionaries𝑫\(ℓ,h\)\\bm\{D\}^\{\(\\ell,h\)\}and \(iv\) the difference\-of\-means steering vector\. The*validation*split selects the steering coefficientα1\\alpha\_\{1\}under the capability gate, and the*test*split is used only for the metrics reported in the tables\.
#### Computational resources
We ran the experiments on one NVIDIA H100 GPU\. We used the HuggingFace Transformers library\([Wolf et al\., 2020](https://arxiv.org/html/2609.31122#bib.bib35)\)to obtain publicly available checkpoints for the Mistral, Llama and Deepseek models\. The total compute time for each experiment is approximately 0\.5 GPU hours\.
### A\.3Method
#### Property dictionary\.
LetpS\(w\)p\_\{S\}\(w\)andpD\(w\)p\_\{D\}\(w\)denote the Dirichlet\-smoothed \(α=0\.01\\alpha\{=\}0\.01\) unigram probabilities of wordwwunder the property\-bearing and non\-property\-bearing fit corpora\. We score each word by its pointwise contribution to the KL divergence,
KLw=pS\(w\)logpS\(w\)pD\(w\),\\mathrm\{KL\}\_\{w\}\\;=\\;p\_\{S\}\(w\)\\log\\frac\{p\_\{S\}\(w\)\}\{p\_\{D\}\(w\)\},\(5\)in both directions, remove stop words, keep the7070highest\-scoring words of each direction and take their union, dropping non\-ASCII entries; this yields139139–140140words per \(model, task\) cell, which we tokenize to obtain𝒫\\mathcal\{P\}and the property subspace𝑼𝒫=𝑾U\[:,𝒫\]\\bm\{U\}\_\{\\mathcal\{P\}\}=\\bm\{W\}\_\{U\}\[:,\\mathcal\{P\}\]\. On roughly 1/3 of the cellsTETthe automatic list is further manually filtered to remove too generic terms\.
#### Steering setting\.
Steering vectors are estimated using the*atonce*strategy: for every selected head, we run a single forward pass over the calibration set with all hooks active and capture activations in parallel, then estimate each head’s difference\-of\-means update on those activations\. This contrasts with an*incremental*strategy, in which one would estimate the update for the first layer’s hook, apply it, then estimate the next layer’s update on the already\-steered activations, and so on\. The atonce variant is faster and avoids hook\-order dependence at the cost of ignoring downstream effects of upstream interventions\.
The steering vector is computed from activations*averaged over the sequence positions*of the prompt\. At inference, the resulting steering vector is applied only to the activation at the*last*token position of the generated continuation\.
### A\.4Evaluation
#### Steering coefficient selection\.
The coefficientα1\\alpha\_\{1\}is the only swept quantity, and it is chosen once per \(method, model, task\) cell on the validation split; the tables report the test run at the selected valueα1∗\\alpha\_\{1\}^\{\*\}\. For all locus variants and DoM the sweep grid is:α1∈\{0\.25,0\.5,0\.75,1,1\.5,…,4\}\\alpha\_\{1\}\\in\\\{0\.25,0\.5,0\.75,1,1\.5,\\dots,4\\\}\. Reference methods are swept with the grid indicated in the respective apperα∈\{5,10,…,30\}\\alpha\\in\\\{5,10,\\dots,30\\\}for ITI, andα∈\{0\.5,…,14\}\\alpha\\in\\\{0\.5,\\dots,14\\\}for SMH\. A grid point is admissible if its validation wiki\-PPL is at most twice the unsteered model’s and its validation MMLU is at least0\.99×0\.99\\timesthe unsteered value;α1∗\\alpha\_\{1\}^\{\*\}is the admissible point with the lowest validation target metric\.
#### Generation\.
OnTETwe sample22completions of6464tokens for each of the615615test prompts \(308308on validation\); onIMDbone completion of5050tokens for each of the1,0001\{,\}000prompts; on Sycophancy22completions of6464tokens for each of the100100\(system prompt, question\) pairs\. Sampling uses temperature1\.01\.0, with top\-pp0\.30\.3and repetition penalty1\.21\.2onTETandIMDb\.
#### Control Metrics\.
PPL\-Wikifollows the protocol of[Rodriguez et al\. \(2025b\)](https://arxiv.org/html/2609.31122#bib.bib31): it is the mean per\-passage perplexity over 20k English Wikipedia articles \(Wikimedia 20231101 dump\), each scored from its opening on the first 50 tokens; validation and test use disjoint halves \(10k passages each\)\. The degeneration gate is relative: an operating point is admissible if its validation PPL\-Wiki is at most twice that of the unsteered model\.MMLUis the 57\-subject 5\-shot accuracy computed with the EleutherAIlm\-evaluation\-harness\([Gao et al\., 2024](https://arxiv.org/html/2609.31122#bib.bib33)\); its 14,042 items are split once, stratified by subject, into a validation half used by the selection gate and a test half reported in the tables \(the Jigsaw and 67B tables use the full set\)\.
#### ITI baseline\.
Following\([Li et al\., 2023](https://arxiv.org/html/2609.31122#bib.bib19)\)\., we train one logistic\-regression probe per attention head \(sklearn LogisticRegression, C=1\.0, L2, max\_\\\_iter=1000, no feature standardization\) on the last\-token per\-head activations \(the o\_\\\_proj inputs\), with classes balanced to the minority count and 2\-fold cross\-validation, and rank heads by mean held\-out accuracy, selecting the top K with K set to LOCuS\-h’s head count\.
#### Where the budgetn=20n\{=\}20comes from\.
Table 6:Choosing the subspace budget from the effective dictionary\.Smallest number of atoms at which the mean fit\-time cumulative EVR over all heads exceeds0\.250\.25\.The value used in the main tables is fixed from the fit data alone, with no access to any steering metric\. For each \(task, model\) cell we replay the fit\-time SOMP decomposition of[Equation3](https://arxiv.org/html/2609.31122#S3.E3)on the effective per\-head dictionaries𝑼~𝒫\(ℓ,h\)\\widetilde\{\\bm\{U\}\}\_\{\\mathcal\{P\}\}^\{\(\\ell,h\)\}with a budget of5050atoms, which yields, for every head, the cumulative EVRe\(ℓ,h\)\(n\)e^\{\(\\ell,h\)\}\(n\)of[Section3](https://arxiv.org/html/2609.31122#S3)as a function of the number of atoms retained\. We then read offn⋆=min\{n:mean\(ℓ,h\)EVR\(ℓ,h\)\(n\)\>0\.25\},n^\{\\star\}\\;=\\;\\min\\Bigl\\\{\\,n\\;:\\;\\operatorname\*\{mean\}\_\{\(\\ell,h\)\}\\mathrm\{EVR\}^\{\(\\ell,h\)\}\(n\)\\;\>\\;0\.25\\,\\Bigr\\\},the smallest budget at which the average head has a quarter of its fit\-time variance reconstructed by its own atoms, with the average taken over all heads\.[Table6](https://arxiv.org/html/2609.31122#A1.T6)reportsn⋆n^\{\\star\}: over the twelve cells it lies in\[14,25\]\[14,25\]\(median2121\), and no cell asks for a budget far from2020, which motivates the single valuen=20n\{=\}20used forLocUSandLocUS\-hs throughout[Table3](https://arxiv.org/html/2609.31122#S4.T3)\. The rule is defined for the per\-head dictionaries; for the per\-layerLocUS\-s we setn=10n\{=\}10by hand\.
## Appendix BHead selection matrices
Figures[3](https://arxiv.org/html/2609.31122#A2.F3),[4](https://arxiv.org/html/2609.31122#A2.F4), and[5](https://arxiv.org/html/2609.31122#A2.F5)visualize the per\-head signals underlyingLocUSand the ITI baseline\. Figure[3](https://arxiv.org/html/2609.31122#A2.F3)plots the EVR matrix over all \(layer, head\) pairs, with heads retained by theμ\+2σ\\mu\+2\\sigmarule marked in orange; Figure[4](https://arxiv.org/html/2609.31122#A2.F4)shows the same scores with heads ranked by EVR within each layer\.
Figure[5](https://arxiv.org/html/2609.31122#A2.F5)shows the linear\-probe validation accuracy used by ITI, in the same layout\. Unlike EVR, high probe accuracy is broadly distributed across heads and layers\.
Figure 3:Per\-head EVR matrices for the nine \(model, task\) cells behind[Table3](https://arxiv.org/html/2609.31122#S4.T3); orange boxes mark theμ\+2σ\\mu\+2\\sigma\-selected heads andkkis their count\.Figure 4:EVR on heads sorted by EVR value\. Per\-head EVR scores for the nine \(model, task\) cells, heads sorted by EVR within each layer \(rank00= highest\)\. Inset reports the number of heads selected by theμ\+2σ\\mu\+2\\sigmarule and their mean EVR\.Figure 5:Sorted probe accuracy for ITI\. Per\-head linear\-probe validation accuracy for the nine \(model, task\) cells, same layout as Figure[4](https://arxiv.org/html/2609.31122#A2.F4)\. Inset reports the head count selected by ITI \(the samekkasLocUS\) and their mean probe accuracy\.
## Appendix CStrength Comparison
[Figure6](https://arxiv.org/html/2609.31122#A3.F6)shows the full validation sweep behind[Tables2](https://arxiv.org/html/2609.31122#S4.T2)and[3](https://arxiv.org/html/2609.31122#S4.T3): for every \(task, model\) cell, the target metric against MMLU asα1\\alpha\_\{1\}grows, for DoM and the fourLocUSvariants, with the selectedα1∗\\alpha\_\{1\}^\{\*\}circled\. On Sycophancy with Llama\-8B no variant, at any strength, moves the score away from the unsteered value \(the whole panel spans less than two points of sycophancy score\), which is why every method reports an unchanged target on that cell\.
Figure 6:Removal–capability trade\-off over the strength sweep, validation split\.Each panel sweeps the steering coefficientα1∈\{0\.25,…,4\}\\alpha\_\{1\}\\in\\\{0\.25,\\dots,4\\\}\(annotated next to each marker\) for one \(task, model\) cell and plots the validation target metric \(xx, lower is better\) against validation MMLU \(yy, higher is better\); the black star is the unsteered model\. Points whose validation wiki\-PPL exceeds twice the unsteered model’s are omitted \(MMLU itself is not gated in the plot\); A black ring marks each variant’s selectedα1∗\\alpha\_\{1\}^\{\*\}, i\.e\. the operating point[Table3](https://arxiv.org/html/2609.31122#S4.T3)reports at test\.Similar Articles
UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering
UniSteer introduces a text-guided activation flow matching method to learn a universal conditional velocity field in activation space, enabling versatile LLM behavior control and classification tasks without task-specific intervention modules.
Multi-Attribute Steering of Language Models via Targeted Intervention
MAT-Steer introduces a novel inference-time intervention framework for steering LLMs across multiple conflicting attributes by learning sparse, orthogonal steering vectors that selectively target tokens relevant to each attribute, achieving gains in QA tasks and generative tasks over prior methods.
GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Models
GeoSteer is an optimization-based method for norm-preserving activation steering in large language models, using geodesic updates on the representation manifold to improve control over model behavior while preserving activation norms.
Deployable Per-Instance Multi-Layer Activation Steering for Large Language Models
This paper introduces a deployable per-instance, multi-layer activation steering technique for large language models, showing that optimal layer selection varies per input and can be predicted from the prompt embedding without gold labels at inference.
Don't Lose Focus: Activation Steering via Key-Orthogonal Projections
This paper introduces Steering via Key-Orthogonal Projections (SKOP), a method to control LLM behavior by preventing attention rerouting, thereby reducing utility degradation while maintaining steering efficacy.