betaFeedback

WorksComputer Science

Right Token, Right Time: From Per-Token Gating to Trajectory-Aware Activation Steering

Abstract

Activation steering via sparse autoencoders (SAEs) enables behavioral control of large language models without task-specific fine-tuning. Yet, standard methods apply the steering signal at every generated token, displacing representations from the distribution the model was trained on and eroding generation quality. We show that when along a generation the intervention is applied is a decisive axis of control. We cast steering as a per-token allocation problem under a fixed budget of steerable tokens, so that methods differ only in which tokens they select, making allocation the sole variable. Within this framework we introduce two gates. The first, Contrastive Feature Projection (CFP), is memoryless and selects the tokens least aligned with the target behavior through a sparse, interpretable feature projection. The second, Closed-Loop Adaptive Bernoulli Steering (CLABS), is trajectory aware. It treats allocation as closed-loop feedback control, tracking the alignment a generation has accumulated so far and spending budget where the trajectory falls behind. On a linear steering-response model we derive an explicit gain region that guarantees closed-loop stability of CLABS and show that its integral term reduces the steady-state tracking offset. Empirically, CFP and CLABS achieve the largest behavioral shift among budgeted gates at matched budgets, with CLABS strongest under the tight budgets where allocation matters most, while best preserving fluency.

Introduction

Large language models (LLMs) increasingly require controllable behavior at inference time, whether suppressing toxicity, adjusting affect, or enforcing stylistic constraints, without the cost of retraining. Activation steering, which adds a fixed direction to a model’s hidden states, has emerged as a lightweight alternative to fine-tuning for such control [39, 33, 42, 11] (Li et al. 2023a). Sparse autoencoders (SAEs) sharpen this approach by decomposing hidden states into interpretable, often monosemantic features [4, 3, 38, 27], which supply behavior-selective directions along which to steer.

Steering acts directly in representation space, and that directness makes it easy to adopt. Most methods apply the intervention uniformly, pushing every generated token by the same vector with the same magnitude [39, 42, 33]. Many tokens, however, already agree with the target behavior, so steering them adds little behavioral change while still displacing the residual stream from the distribution the model was trained on [24] (Li et al. 2023a). The displacement accumulates across the sequence, and uniform steering therefore trades away fluency as behavioral strength grows, a degradation reported whenever intervention magnitude is pushed high [33, 38] (Li et al. 2023a) and one we reproduce. If a smaller set of tokens carries most of the behavioral signal, effective control hinges on when to intervene, a question uniform steering never asks.

Recent work has begun to ask it from several directions. Adaptive magnitude control regulates how much to steer with a feedback controller across layers, but still intervenes at every position [31]. Stochastic sparse steering meets a budget by flipping an independent coin per token [11], so its selection is unrelated to need. Conditional steering intervenes only where a trigger fires, for instance on tokens that activate SAE toxicity features [18], at high-entropy positions selected with a learned value model and multi-sample search (Hu et al. 2026), or when tracked states indicate that steering is necessary, with backtracking to regenerate offending spans [16]. None of these, however, combines two properties that a fixed-budget gate needs. The first is calibration to an explicit intervention budget, so that methods can be compared at equal cost. The second is a firing rule that responds to the alignment the generation has accumulated so far, rather than to the current token alone, without auxiliary rollouts.

We study this question within one framework, sparse token-level steering, in which a gate decides whether to intervene subject to a budget pp, the fraction of tokens that may be steered. Treating steering as allocation, that is, which tokens to steer rather than how hard to push, turns every comparison into a controlled ablation, since at a fixed budget methods differ only in which tokens they select. We instantiate the framework with two gates. Contrastive Feature Projection (CFP) reads each token’s instantaneous alignment along a sparse, interpretable feature direction and spends the budget on the least-aligned positions, at the cost of one dot product per token. CFP is informed but memoryless, so it cannot distinguish a momentary dip inside an on-target span from a sustained drift away from the target. Closed-Loop Adaptive Bernoulli Steering (CLABS), our central method, adds this memory. It treats allocation as feedback control, setting a budget-calibrated steering probability from the current misalignment and a leaky integral of recent misalignment, so that sustained drift draws more budget than isolated dips. CLABS runs within one decoding trajectory, without auxiliary rollouts, backtracking, or a learned scorer. Our contributions are as follows.

  • We cast token-level activation steering as a budgeted allocation problem, in which when to spend a fixed intervention budget is the single decision variable, and introduce two efficient activation-space gates, the memoryless CFP and the trajectory-aware CLABS.

  • We analyze CLABS as a proportional–integral controller with a leaky integral term. Under a linear approximation of the steering response, common in feedback analysis [31], we derive an explicit gain region that guarantees closed-loop stability and bounded tracking error, and show that integral action reduces, but does not remove, the steady-state offset. To our knowledge, this is the first stability certificate for a sparse steering gate.

  • On two base models and three span-level targets, CFP and CLABS achieve the largest behavioral shift among budgeted gates at matched realized budgets, with the widest margin at tight budgets. Among the budgeted gates they best preserve fluency and diversity, and under toxicity reduction they are the only gates preferred to the unsteered model on instruction following at every budget under both judges. A component ablation isolates the contribution of the history term.

Related Work

Activation steering. Activation steering adds a fixed direction to a model’s hidden states at inference time, controlling behavior without weight updates [36, 39, 33, 42] (Li et al. 2023a). Directions have been derived from class-conditional mean differences [30], in-context demonstrations [29], weight arithmetic [15], and expert (de)activation in Mixture-of-Experts models [12]. Sparse autoencoders decompose the residual stream into sparse, near-monosemantic features [4, 3, 38, 27], whose decoder rows provide interpretable steering directions [34]. AxBench (Wu et al. 2025) shows that the choice of direction matters and that simple baselines can rival SAE-based ones, which is why we hold the direction fixed across all methods and vary only allocation. Steering now spans detoxification, truthfulness, hallucination reduction, style, safety, and instruction following [39, 41, 6, 35] (Li et al. 2023a), complementing decoding-time detoxification controllers [28] (Dhathri et al. 2020) evaluated on RealToxicityPrompts [17]. Most of these methods apply the direction at every generated token, and this dense intervention degrades fluency as behavioral shift grows [39, 33] (Wu et al. 2025). Recent work accordingly evaluates steering on the trade-off between behavioral strength and quality rather than on either alone [14], a view we adopt by reporting efficacy and quality together at matched budgets.

Sparse, conditional, and adaptive steering. Recent work relaxes the every-token assumption along different axes. AUSTEER [13] sparsifies which dimensions are edited but still fires at every position, and is orthogonal to the temporal sparsity we study. Stochastic Token Steering (STS; [11]) gates each token with an independent Bernoulli(pp) draw, which meets a budget but selects positions independently of the state. Conditional methods intervene only where a trigger fires. Breaking Bad Tokens [18] applies SAE-derived interventions at tokens that activate toxicity features, a threshold rule that corresponds to CFP in our framework once calibrated to a budget. Sparse Inference-time Alignment (SIA; Hu et al. 2026) steers high-entropy junctions in logit space through a learned, reward-distilled value model coupled to Best-of-NN search. FASB [16] tracks internal states over generated tokens to decide whether and how strongly to steer, and backtracks to regenerate tokens when steering is deemed necessary. These methods are governed by thresholds, rewards, or regeneration rather than by an intervention budget, and their firing decisions do not feed back on the alignment accumulated along a single trajectory. Where a single-pass selection signal can be isolated, as for the feature trigger and the entropy criterion, we compare against it as a budget-calibrated gate rather than against the complete system. PID Steering [31] also draws on classical control, but regulates the steering magnitude at every token with an integral term accumulated across layers and admits no budget. It acts on a complementary axis, since a gate decides when to steer and a magnitude controller how much, and the two compose. CLABS holds the magnitude fixed and applies PI feedback along the token sequence to set a budget-calibrated gating probability. To our knowledge, it is the first gate that allocates a fixed steering budget by feedback on accumulated alignment within one decoding trajectory, without rollouts or backtracking, and the first sparse steering gate with a closed-loop stability analysis.

Method

We present two gates that spend the same budget in different ways. Both read a single alignment signal, the projection of the residual stream onto a steering direction built from sparse autoencoder features. CFP is memoryless and steers the least-aligned tokens. CLABS feeds the same signal into a proportional–integral (PI) controller with a leaky integral term, so its steering probability also reflects the alignment the generation has accumulated so far. We then analyze the resulting closed loop. Because both gates are calibrated to the same budget, allocation is the only difference between them.

Notation and Setup

Let pθp_\theta be an autoregressive language model and htℓ∈Rdh_t^\ell\in\mathbb{R}^d the residual-stream activation at a fixed intervention layer ℓ\ell and decoding step tt. We write ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle for the Euclidean inner product, ∥⋅∥\|\cdot\| for the induced norm, 1[⋅]\mathbb{1}[\cdot] for the indicator, sign⁡(⋅)\operatorname{sign}(\cdot) for the sign, μ\mu, σ\sigma for a sample mean and standard deviation, ϵ>0\epsilon> 0 for a small constant, and clip⁡(x;a,b)=min⁡(max⁡(x,a),b)\operatorname{clip}(x;a,b)=\min(\max(x,a),b).

A token-level steering intervention adds a fixed direction to the residual stream, gated per token,

h~tℓ=htℓ+mtαv,(1)\tilde{h}_t^\ell= h_t^\ell+ m_t\alpha v, \tag*{(1)}

where v∈Rdv \in\mathbb{R}^d with ∥v∥=1\|v\|=1 is the steering direction, α∈R\alpha\in\mathbb{R} the strength, and mt∈{0,1}m_t \in\{0,1\} the gate. This separates two control axes, allocation, given by the gate sequence m1:Tm_{1:T}, and magnitude, given by α\alpha. We hold α\alpha fixed and study allocation. Every method runs at the same budget p∈(0,1)p \in(0,1), the expected fraction of steered tokens. Differences in downstream behavior are therefore attributable to which tokens a method steers.

Steering Direction from Sparse Features

We derive the direction from a sparse autoencoder (SAE) rather than from raw activation differences. SAEs decompose the residual stream into sparse, approximately monosemantic features [4, 3]. A direction built from a few behavior-selective features is therefore expected to be less entangled with unrelated attributes than a dense difference-in-means vector, which matters when the goal is to move one behavior while leaving generation quality intact.

Let the SAE at layer ℓ\ell have encoder f:Rd→RFf:\mathbb{R}^d\to\mathbb{R}^F, decoder Wdec∈RF×dW_{\mathrm{dec}}\in\mathbb{R}^{F\times d}, and reconstruction SAE⁡(h)=Wdec⊤f(h)\operatorname{SAE}(h)=W_{\mathrm{dec}}^\top f(h). Let D+\mathcal{D}^{+} be a target corpus that exhibits the behavior of interest, namely the behavior to be induced for elicitation or the behavior to be removed for suppression, and let D−\mathcal{D}^{-} be a neutral corpus. We compute the per-feature differential

Δi=1∣D+∣∑x∈D+fi(x)−1∣D−∣∑x∈D−fi(x),(2)\Delta_i=\frac{1}{|\mathcal{D}^{+}|}\sum_{x\in\mathcal{D}^{+}}f_i(x)-\frac{1}{|\mathcal{D}^{-}|}\sum_{x\in\mathcal{D}^{-}}f_i(x), \tag*{(2)}

where fi(x)f_i(x) denotes the ii-th SAE code averaged over the tokens of xx. We take the indices I\mathcal{I} of the top-kk features by Δi\Delta_i and form the direction from the corresponding decoder rows,

v=∑i∈IWdec,i∥∑i∈IWdec,i∥.(3)v=\frac{\sum_{i\in\mathcal{I}}W_{\mathrm{dec},i}}{\left\|\sum_{i\in\mathcal{I}}W_{\mathrm{dec},i}\right\|}. \tag*{(3)}

Since vv points toward D+\mathcal{D}^{+}, suppression tasks use α<0\alpha<0 to move activations away from it, while elicitation tasks use α>0\alpha>0. We define a task-agnostic behavioral probe

u=sign⁡(α) v,(4)u=\operatorname{sign}(\alpha)\,v, \tag*{(4)}

so that a larger ⟨htℓ,u⟩\langle h_t^\ell,u\rangle always indicates that htℓh_t^\ell is more aligned with the desired behavior, regardless of the task sign. The probe is a fixed vector, so reading it requires no SAE forward pass at inference time.

CFP: Contrastive Feature Projection

A token already aligned with the target has little to gain from intervention, whereas a token far from the target is where a fixed-strength push is expected to change the trajectory most. CFP reads each token’s instantaneous alignment along the probe and spends the budget on the least-aligned positions,

stCFP=⟨htℓ,u⟩,mt=1[stCFP≤τ],(5)s_t^{\mathrm{CFP}}=\langle h_t^\ell,u\rangle,\qquad m_t=\mathbb{1}[s_t^{\mathrm{CFP}}\leq\tau], \tag*{(5)}

with τ\tau calibrated to the budget pp. By sparing tokens that are already on target, CFP avoids perturbing the residual stream where steering would add little. CFP is the budget-calibrated form of conditional, feature-triggered steering. Given τ\tau, each mtm_t depends only on the current activation, so CFP cannot tell a momentary dip inside an on-target span from a sustained drift away from the target. It may therefore keep steering an aligned span, or stop too early on one that is not. This limitation motivates CLABS.

CLABS: Closed-Loop Adaptive Bernoulli Steering

CLABS closes the loop. Instead of thresholding a static signal, it treats allocation as discrete-time feedback control [2] on the generation’s running alignment, so the budget is spent in response to what has already been generated. We first standardize the projection with calibration statistics μ,σ\mu,\sigma,

s^t=⟨htℓ,u⟩−μσ+ϵ,(6)\hat{s}_t=\frac{\langle h_t^\ell,u\rangle-\mu}{\sigma+\epsilon}, \tag*{(6)}

which expresses alignment in standard deviations above or below its typical value. We then choose a setpoint s∗s^* in the same units. It is the internal alignment level CLABS aims to hold, with s∗=0s^*=0 targeting the calibration mean and larger s∗s^* demanding stronger alignment. Given s∗s^*, we define the tracking error and its leaky integral, an exponentially weighted moving average (EWMA),

et=s∗−s^t,(7)e_t=s^*-\hat{s}_t, \tag*{(7)}
eˉt=(1−β)eˉt−1+βet,eˉ0=0,(8)\bar{e}_t=(1-\beta)\bar{e}_{t-1}+\beta e_t,\qquad\bar{e}_0=0, \tag*{(8)}

where β∈(0,1]\beta\in(0,1] sets the controller’s memory, so that eˉt\bar{e}_t summarizes the deviation accumulated over roughly the last 1/β1/\beta tokens. CLABS then sets a PI steering probability and samples the gate,

pt=clip⁡(p0+κpet+κieˉt;0,1),(9)p_t=\operatorname{clip}(p_0+\kappa_p e_t+\kappa_i\bar{e}_t;0,1), \tag*{(9)}
mt∼Bernoulli⁡(pt),(10)m_t\sim\operatorname{Bernoulli}(p_t), \tag*{(10)}

with bias p0p_0 and gains κp,κi≥0\kappa_p,\kappa_i\geq0. The proportional term κpet\kappa_p e_t reacts to the current token’s misalignment, the same information CFP uses. The leaky integral term κieˉt\kappa_i\bar{e}_t adds what memoryless gates lack, a record of recent tokens have stayed behind the setpoint. The steering probability therefore rises while the trajectory lags and relaxes once it catches up, so a sustained drift draws more budget than an isolated dip. The bias p0p_0 is calibrated so that the realized gate rate matches the target budget pp, as detailed in Appendix B. Algorithm 1 summarizes the procedure. Unlike

Algorithm 1: CLABS

Require: probe uu, direction vv, strength α\alpha, setpoint s⋆s^\star, gains κp,κi,β\kappa_p,\kappa_i,\beta, bias p0p_0, stats μ,σ\mu,\sigma

  1. eˉ←0\bar{e} \leftarrow0

  2. for t=1,…,Tt=1,\ldots,T do

  3. obtain residual htℓh_t^\ell at layer ℓ\ell

  4. s^t←(⟨htℓ,u⟩−μ)/(σ+ϵ)\hat{s}_t \leftarrow(\langle h_t^\ell,u\rangle-\mu)/(\sigma+\epsilon)

  5. et←s⋆−s^te_t \leftarrow s^\star-\hat{s}_t; eˉ←(1−β)eˉ+βet\bar{e}\leftarrow(1-\beta)\bar{e}+\beta e_t

  6. pt←clip⁡(p0+κpet+κieˉ;0,1)p_t\leftarrow\operatorname{clip}(p_0+\kappa_p e_t+\kappa_i\bar{e};0,1)

  7. sample mt∼Bernoulli⁡(pt)m_t\sim\operatorname{Bernoulli}(p_t)

  8. if mt=1m_t=1 then

  9. h~tℓ←htℓ+αv\tilde{h}_t^\ell\leftarrow h_t^\ell+\alpha v, then apply norm preservation and projection clamping

  10. end if

  11. decode the next token from the (possibly steered) state

  12. end for

conditional methods that rely on a learned scorer with multi-sample search (Hu et al. 2026) or on backtracking [16], CLABS needs none of these. To our knowledge, it is the first gate that sets a budget-calibrated steering probability by feedback on accumulated alignment within a single decoding trajectory.

Theoretical Analysis of CLABS

A gate driven by the model’s own state forms a feedback loop. Steering changes alignment, alignment changes the next steering decision, and loops of this kind can oscillate or drift. Before evaluating CLABS empirically, we therefore ask whether its controller is stable and which gains keep it stable. We answer both questions on a first-order steering-response model of how steering moves alignment, following standard practice in feedback control of computing systems [20, 2] and prior control-theoretic steering [31]. The model describes the interior operating point, where the calibrated budget keeps ptp_t strictly inside (0,1)(0,1). The result is therefore a stability certificate for this linear approximation, not a worst-case guarantee for the full nonlinear decoding process. Behavioral efficacy at matched budgets is assessed separately in § 4.

Assumption 1 (Linear steering response). Let Ft−1\mathcal{F}_{t-1} denote the generation history available when ptp_t is computed. In the interior regime there exist constants ϕ∈[0,1)\phi\in[0,1), g>0g>0, and β0∈R\beta_0\in\mathbb{R} such that

s^t+1=ϕs^t+(1−ϕ)β0+gmt+ηt.(11)\hat{s}_{t+1}=\phi\hat{s}_t+(1-\phi)\beta_0+g m_t+\eta_t. \tag*{(11)}

Here ϕ\phi is the activation persistence, measuring how strongly alignment carries over from one token to the next, gg is the steering gain, measuring how much steering a token shifts next-token alignment, and β0\beta_0 is the unsteered equilibrium. The disturbance ηt\eta_t has zero mean given Ft−1\mathcal{F}_{t-1} and conditional variance at most σ2\sigma^2, and it absorbs token-level fluctuations and neglected higher-order effects.

In the interior regime, where the clip in Eq. (9) is inactive, combining Eq. (11) with the PI law and the leaky integral of

(8) shows that the controller state xt=(et,eˉt)Tx_t=(e_t,\bar{e}_t)^\mathsf{T} follows the linear recursion

xt+1=Axt+b+wt,A=(a−gκiβa(1−β)−βgκi),(12)x_{t+1}=Ax_t+b+w_t,\quad A=\begin{pmatrix}a&-g\kappa_i\\\beta a&(1-\beta)-\beta g\kappa_i\end{pmatrix}, \tag*{(12)}

where a=ϕ−gκpa=\phi-g\kappa_p, bb is a constant drift, and wt=−rη~tw_t=-r\tilde{\eta}_t with r=(1,β)Tr=(1,\beta)^\mathsf{T} is a zero-mean disturbance. The effective noise η~t\tilde{\eta}_t accounts for both the model disturbance ηt\eta_t and the randomness of the Bernoulli gate, and its conditional variance is bounded by a constant σˉ2\bar{\sigma}^2. The derivation of this recursion, the explicit form of bb, and the bound σˉ2\bar{\sigma}^2 are given in Appendix A. The matrix AA couples how the model responds to steering (ϕ,g)(\phi,g) with how the controller reacts (κp,κi,β)(\kappa_p,\kappa_i,\beta), so stability is a joint property of the model and the gate.

Theorem 1 (Stability and tracking). Under Assumption 1, let ρ(A)\rho(A) be the spectral radius of AA in Eq. (12), II the 2×22\times2 identity, and [z]1[z]_1 the first entry of a vector zz. A sufficient condition for ρ(A)<1\rho(A)<1 is

0≤a<1,0≤gκi<(2−β)(1+a)β.(13)0\leq a<1,\qquad0\leq g\kappa_i<\frac{(2-\beta)(1+a)}{\beta}. \tag*{(13)}

Under this condition the following hold.

  1. E[xt]\mathbb{E}[x_t] converges geometrically, with rate arbitrarily close to ρ(A)\rho(A), to the steady-state controller state xss=(I−A)−1bx_{\mathrm{ss}}=(I-A)^{-1}b.

  2. lim sup⁡t→∞Var⁡(et)≤c(A)σˉ2\limsup_{t\to\infty}\operatorname{Var}(e_t)\leq c(A)\bar{\sigma}^2, where c(A)=∑k≥0[(Akr)1]2≥1c(A)=\sum_{k\geq0}[(A^k r)_1]^2\geq1 is a gain-dependent variance multiplier, finite whenever ρ(A)<1\rho(A)<1.

  3. With ess=[xss]1e_{\mathrm{ss}}=[x_{\mathrm{ss}}]_1 the steady-state tracking error,

1T∑t=1TE[(et−ess)2]≤O(1T)+O(σˉ2).(14)\frac{1}{T}\sum_{t=1}^{T}\mathbb{E}\left[(e_t-e_{\mathrm{ss}})^2\right]\leq O\left(\frac{1}{T}\right)+O(\bar{\sigma}^2). \tag*{(14)}

The proof is given in Appendix A. The theorem reduces stability of the gated loop to a closed-form condition that involves only ϕ\phi, gg, and the controller gains, so any candidate (κp,κi,β)(\kappa_p,\kappa_i,\beta) can be checked against Eq. (13) given estimates of ϕ\phi and gg. On the linear model, a choice inside the box guarantees that the mean controller state converges geometrically, that the asymptotic variance of the tracking error is at most c(A)σˉ2c(A)\bar{\sigma}^2, and that the transient contribution to the time-averaged squared error vanishes at rate O(1/T)O(1/T). To our knowledge, no prior sparse steering gate comes with such a certificate.

Corollary 1 (Integral action reduces the steady-state offset). Under condition (13), with ϕ\phi, gg, κp\kappa_p, κi\kappa_i, β\beta and the model constants ϕ,g,β0\phi,g,\beta_0 fixed, the deterministic steady-state error is

ess=N1−ϕ+g(κp+κi),N=(1−ϕ)(s⋆−β0)−gp0.(15)e_{\mathrm{ss}}=\frac{N}{1-\phi+g(\kappa_p+\kappa_i)},\qquad N=(1-\phi)(s^\star-\beta_0)-g p_0. \tag*{(15)}

Hence, as κi\kappa_i varies within the box, ∣ess∣|e_{\mathrm{ss}}| is non-increasing in κi\kappa_i, and strictly decreasing whenever N≠0N\ne0.

The proof is given in Appendix A.

Remark 1 *The two inequalities in Eq. (13) correspond to the two controller terms. The first keeps the proportional correction gκpg\kappa_p no larger than the persistence ϕ\phi, so the proportional response alone never flips the sign of the error. The second caps the integral gain by a bound that grows with the memory length 1/β1/\beta, so slow integration is easier to stabilize but reacts more sluggishly. The same box also limits how much integral action can help. Here NN is (1−ϕ)(1-\phi) times the gap between the setpoint and the alignment an open-loop gate with bias p0p_0 would reach. Combining Eq. (15) with Eq. (13) gives, for every admissible gain at a fixed p0p_0 and N≠0N \ne0,

∣ess∣>∣N∣1−a+(2−β)(1+a)/β>0.(16)\lvert e_{\mathrm{ss}}\rvert> \frac{\lvert N\rvert}{1-a+(2-\beta)(1+a)/\beta} > 0. \tag*{(16)}

Unlike a classical integrator, the leaky integral therefore trades exact tracking for bounded memory, reducing the offset. The derivation is given in Appendix § 6.

Experiments

Our experiments test two claims. The first is that at a fixed intervention budget the choice of which tokens are steered materially changes the behavioral outcome. The second is that trajectory-aware allocation improves on memoryless allocation, and that the improvement is largest where the budget is tight. Within each model and target, every method shares the steering direction, intervention layer, per-token strength, decoding settings, and post-intervention corrections, and differs only in when the direction is injected. The gating policy is therefore the sole independent variable.

Setup

We evaluate two open base models that differ in scale and family, Gemma 2 2B (Team et al. [37]) and Llama 3.1 8B (Dubey et al. [8]). Following established practice, we intervene on the residual stream at a single mid-network layer (Labonne [23]), and read both the steering direction and the gating signals from a pretrained residual-stream SAE at that layer. We study two control polarities over three span-level targets. Suppression uses RealToxicityPrompts (Gehman et al. [17]), whose continuations we detoxify. Elicitation uses GoEmotions (Demszky et al. [7]), collapsed to the Ekman categories (Ekman [10]), where we induce fear or sadness from neutral prompts. To verify that control does not come at the expense of competence, we also evaluate instruction following on AlpacaEval ([9] (Li et al. 2023b)). All results are averaged over three random seeds. Three lightweight corrections follow Eq. (1), namely norm preservation (Turner et al. [39]), activation clamping (Labonne [23]), and a decoding-stage repetition penalty (Keskar et al. [22]). They are applied identically to every method, so no gate receives a correction the others do not. Further implementation details are given in Appendix § 7.

Baselines

All methods share the steering intervention and differ only in the binary gate mt∈{0,1}m_t \in\{0,1\} that decides when the shared direction is injected. We compare against three unbudgeted reference points that bound the achievable range and four budgeted gates that, like ours, allocate a fixed budget pp.

  1. No-steer sets mt=0m_t = 0 and serves as the uncontrolled reference against which behavioral shift is measured.

  2. Full-steer sets mt=1m_t = 1, the dense fixed-magnitude limit of our framework.

  3. Prompting prepends a task-specific safety or emotion instruction, refined by iterative prompt engineering, to the input (Appendix § 9).

Adaptive magnitude controllers such as PID Steering (Nguyen et al. [31]) operate on a complementary axis. They modulate how much to steer at every position rather than when to steer, and therefore admit no budget p<1p < 1. A gate and a magnitude controller compose naturally, so we do not treat them as competitors under matched allocation.

Two budgeted gates are derived from Sparse Inference-time Alignment (SIA) (Hu et al. 2026). SIA is governed by a reward threshold rather than a token budget and relies on a learned value model and multi-sample search, so it cannot be run at a matched budget as published. We therefore extract its selection signal, the predictive entropy of the next-token distribution, and turn it into a single-pass gate calibrated to the same pp. This isolates SIA’s selection hypothesis, that intervention should target tokens where the model is uncertain, from its scorer and search, and is not a reproduction of the full method.

  1. SIA-PE (Hu et al. 2026) fires on tokens of high predictive entropy.

  1. SIA-ER (Hu et al. 2026) fires on a relative rise in entropy from the previous token.

  1. SAE Reconstruction Error (SRE), which we introduce as a mechanistic counterpoint to the uncertainty gates, steers the tokens that the dictionary represents faithfully, mt=1[∥ht−SAE⁡(ht)∥∥ht∥+ϵ≤τ]m_t = \mathbb{1}\left[\frac{\lVert h_t-\operatorname{SAE}(h_t)\rVert}{\lVert h_t\rVert+\epsilon}\le\tau\right].

  1. STS (random) (Eshragh et al. 2026) steers each token independently with mt∼Bernoulli⁡(p)m_t \sim\operatorname{Bernoulli}(p). Its selection is unrelated to any token property, so it isolates the effect of the budget alone and serves as the allocation-agnostic control.

Because these gates respond to different signals, a shared threshold would yield different budgets. We therefore calibrate every gate to the same target pp on a held-out split by tuning its single control parameter under active steering (Appendix § 7). Matching the budget at calibration time does not by itself guarantee matched spending at test time, so we also measure the realized intervention rate on the evaluation prompts. Realized rates stay close to the target for every method, and CFP and CLABS do not spend more than the gates they outperform (Appendix § 8).

Evaluation and Metrics

We assess each method along two axes, behavioral efficacy and the quality cost at which it is obtained. Efficacy is measured by the behavioral shift, the change in an external classifier’s score on the continuation relative to No-steer. The prompt is excluded, and the shift is signed by task polarity so that a larger value always means a larger move toward the target. Toxicity is scored with a RoBERTa classifier and emotion with a DistilRoBERTa classifier fine-tuned on the Ekman categories (Hartmann et al. [19]). Further details are reported in Appendix § 7.

Utility is measured by the AlpacaEval win rate against No-steer under two independent judges, gpt-5-mini and Claude Sonnet 4.6 [32, 1], with position randomization. A value above 50 means the steered response is preferred. We complement the win rate with perplexity, rep-3, and distinct-2, which track fluency, degeneration, and diversity of the same responses.

Results

Behavioral efficacy. Figure 1 reports the behavioral shift on all six model–target panels, with shaded bands denoting the standard deviation. At matched budgets CFP and CLABS achieve the largest shift among the budgeted gates on every panel. Their margin is widest at the smallest budgets, where a scarce allowance must be spent well, and there they reach two to three times the shift of the uncertainty gates and the random control while steering the same number of tokens. Which tokens are steered, and not only how many, therefore governs efficacy. As the budget grows the gates converge, as expected when nearly every token may be steered. Already at a moderate budget CLABS recovers 89–92% of the Full-steer shift on the emotion targets, so steering about half of the tokens attains most of the effect of steering all of them. Both of our gates also surpass Prompting, whose shift stays fixed, once the budget is no longer minimal.

Figure 1. Behavioral shift versus intervention ratio pp on two models and three targets. CFP and CLABS achieve the largest shift among the budgeted gates, with the widest margin at low pp.

The lead of CLABS at small budgets follows from its design. The integral term accumulates the running misalignment of the generation, raising the steering probability while a span drifts off target and relaxing it once alignment is restored, so the budget concentrates on the spans that actually lag. A memoryless gate cannot allocate this way and may spend budget on a token that dips briefly inside a span already on target, and the uncertainty gates fire on entropy that is only loosely tied to the target behavior, which is why SIA-PE and SIA-ER fall furthest behind when the budget is tight. The pattern holds on both a 2B and an 8B model from different families, and the ranking is unchanged when methods are compared at their realized rather than nominal budgets (Appendix § 8).

Instruction following. Behavioral control is useful only if the model stays helpful. We steer Gemma 2 2B with the toxicity-reduction direction on AlpacaEval instructions and compare each steered response with its No-steer counterpart (Table 1). CFP and CLABS are the only gates preferred to No-steer at every budget under both judges. Their fluency, repetition, and diversity degrades the least, whereas the other gates lose quality steadily as pp grows and Full-steer degrades markedly. The two judges differ in spread but agree on this ranking.

MethodppGPT-5-miniSonnet 4.6PPL ↓\downarrowrep-3 ↓\downarrowdistinct-2 ↑\uparrow
No-Steer0––23.0±2.423.0 \pm2.40.01±0.0040.01 \pm0.0040.91±0.03\mathbf{0.91 \pm0.03}
Prompting–42.4±2.642.4 \pm2.646.3±2.646.3 \pm2.623.5±2.523.5 \pm2.50.01±0.0050.01 \pm0.0050.90±0.030.90 \pm0.03
Full-Steer1.043.2±2.943.2 \pm2.944.9±2.844.9 \pm2.841.2±6.841.2 \pm6.80.09±0.030.09 \pm0.030.68±0.070.68 \pm0.07
SIA-ER0.252.1±2.752.1 \pm2.756.7±2.656.7 \pm2.625.1±2.725.1 \pm2.70.02±0.0070.02 \pm0.0070.88±0.040.88 \pm0.04
SIA-ER0.550.8±2.850.8 \pm2.855.0±2.755.0 \pm2.729.4±3.629.4 \pm3.60.04±0.010.04 \pm0.010.82±0.050.82 \pm0.05
SIA-ER0.844.3±3.144.3 \pm3.149.5±2.949.5 \pm2.938.7±6.138.7 \pm6.10.08±0.030.08 \pm0.030.72±0.070.72 \pm0.07
SIA-PE0.258.6±2.558.6 \pm2.556.7±2.556.7 \pm2.524.8±2.624.8 \pm2.60.02±0.0060.02 \pm0.0060.89±0.040.89 \pm0.04
SIA-PE0.554.0±2.754.0 \pm2.754.0±2.754.0 \pm2.728.1±3.328.1 \pm3.30.03±0.010.03 \pm0.010.84±0.050.84 \pm0.05
SIA-PE0.847.1±3.047.1 \pm3.049.2±2.949.2 \pm2.936.2±5.736.2 \pm5.70.07±0.020.07 \pm0.020.74±0.060.74 \pm0.06
SRE0.251.2±2.651.2 \pm2.653.5±2.653.5 \pm2.624.3±2.624.3 \pm2.60.02±0.0060.02 \pm0.0060.89±0.040.89 \pm0.04
SRE0.553.4±2.653.4 \pm2.654.7±2.654.7 \pm2.627.0±3.127.0 \pm3.10.03±0.010.03 \pm0.010.85±0.050.85 \pm0.05
SRE0.849.0±2.949.0 \pm2.949.6±2.849.6 \pm2.833.5±5.033.5 \pm5.00.06±0.020.06 \pm0.020.77±0.060.77 \pm0.06
STS (random)0.251.0±2.751.0 \pm2.753.2±2.753.2 \pm2.725.6±2.825.6 \pm2.80.02±0.0070.02 \pm0.0070.88±0.040.88 \pm0.04
STS (random)0.550.1±2.850.1 \pm2.852.3±2.852.3 \pm2.830.2±3.830.2 \pm3.80.04±0.010.04 \pm0.010.81±0.050.81 \pm0.05
STS (random)0.843.5±3.243.5 \pm3.248.5±3.048.5 \pm3.039.9±6.439.9 \pm6.40.09±0.030.09 \pm0.030.70±0.070.70 \pm0.07
CFP (ours)0.253.0±2.653.0 \pm2.656.4±2.556.4 \pm2.523.1±2.423.1 \pm2.40.01±0.0040.01 \pm0.0040.90±0.030.90 \pm0.03
CFP (ours)0.557.1±2.557.1 \pm2.556.4±2.556.4 \pm2.524.5±2.624.5 \pm2.60.02±0.0060.02 \pm0.0060.89±0.040.89 \pm0.04
CFP (ours)0.852.2±2.752.2 \pm2.753.8±2.653.8 \pm2.627.8±3.227.8 \pm3.20.03±0.010.03 \pm0.010.86±0.040.86 \pm0.04
CLABS (ours)0.252.5±2.652.5 \pm2.656.4±2.556.4 \pm2.522.8±2.322.8 \pm2.30.01±0.0030.01 \pm0.0030.91±0.030.91 \pm0.03
CLABS (ours)0.556.2±2.556.2 \pm2.555.6±2.555.6 \pm2.523.9±2.523.9 \pm2.50.01±0.0040.01 \pm0.0040.90±0.030.90 \pm0.03
CLABS (ours)0.851.0±2.851.0 \pm2.852.6±2.652.6 \pm2.626.5±3.026.5 \pm3.00.02±0.0060.02 \pm0.0060.88±0.040.88 \pm0.04

Table 1. Table 1: Instruction following and generation quality on AlpacaEval under toxicity-reduction steering. Win rate is the fraction of instructions where the steered response beats No-steer. PPL, rep-3, and distinct-2 track fluency and diversity. CFP and CLABS retain utility as pp grows, while the other gates and Full-steer lose quality.

A single high win-rate cell should be read together with efficacy. SIA-PE attains the highest gpt-5-mini win rate at the smallest budget, yet its behavioral shift in that regime is among the weakest (Figure 1), its quality erodes quickly as pp grows, and Sonnet scores it slightly above our gates. This reflects a light touch rather than effective control. The relevant criterion is to sustain efficacy and utility jointly across budgets and judges, and among the gating rules we compare only CFP and CLABS meet it. Because every gate shares the same direction, strength, and corrections, this difference is attributable to allocation. Sparse, well-placed steering leaves most of the trajectory unperturbed, whereas STS spends its budget irrespective of need and Full-steer gives up temporal sparsity altogether.

Further results, including realized budgets, ablations, and additional instruction-following settings, are given in Appendix C, and limitations are discussed in Appendix E.

Discussion

Why allocation and history matter. If every steering step shifted alignment equally, any allocation of a fixed budget would reach the same mean behavior. Allocation matters because the effect of steering varies across tokens, as the consistent gap between the informed gates and random gating at matched realized budgets shows. CFP and CLABS read the same alignment signal, so the gap between them reflects the value of memory. The leaky integral weights persistent errors more than fluctuating ones, so CLABS spends more of its budget on sustained drift than on momentary dips, which a memoryless gate treats alike. This should matter most when the budget is scarce, where CLABS leads in Figure 1.

Practical guidance. Neither gate needs an SAE pass, an extra forward pass, or multi-sample decoding. CFP adds one inner product per token and CLABS a few scalar operations on top of it, which is on the order of 10−610^{-6} of the cost of a forward pass and is therefore negligible. CLABS is preferable under tight budgets, and CFP is a simpler, competitive choice at larger budgets, where the informed gates converge.

Conclusion

We framed activation steering as budgeted allocation, in which a per-token gate decides when to spend a fixed intervention budget, so that allocation is the sole variable. We introduced two gates. CFP is informed but memoryless and steers the least-aligned tokens by an interpretable feature signal. CLABS is trajectory aware and spends budget where a generation lags. On a linear steering-response model, its PI controller is provably stable within an explicit gain region, and its integral term reduces the steady-state offset. Across two models and three targets, both gates achieve the largest behavioral shift among budgeted gates, CLABS most clearly at the tight budgets where allocation matters most. They also best preserve fluency compared to the budgeted gates and full-steering.

Content Warning. This paper studies inference-time control of language model behavior, including the reduction of toxic generation. The supplementary material contains examples of toxic, offensive, and emotionally distressing model outputs that are shown to illustrate the method. Reader discretion is advised.

Theoretical Results

This appendix gives the derivations and proofs behind § 3.5. We first derive the closed-loop recursion, including the noise contributed by the Bernoulli gate (Section A.1). We then prove the stability condition of Theorem 1 (Section A.2), its three claims (Section A.3), Corollary 1, and the offset bound of Remark 1 (Section A.4). Finally, we present the error weighting that underlies the role of history in CLABS (Section A.5).

All guarantees concern the linear steering-response model of Assumption 1 in the interior regime, where the clip in Eq. (9) is inactive. They certify that the gated loop is stable on this model. They are not worst-case guarantees for the full nonlinear decoding process, and the efficacy of CLABS at matched budgets is assessed empirically.

Notation. We recall the quantities of Section 3. The standardized alignment is s^t\hat{s}_t, the setpoint is s⋆s^\star, the tracking error is et=s⋆−s^te_t = s^\star-\hat{s}_t, and its leaky integral is eˉt=(1−β)eˉt−1+βet\bar{e}_t = (1-\beta)\bar{e}_{t-1}+\beta e_t with eˉ0=0\bar{e}_0 = 0. In the interior regime the steering probability is pt=p0+κpet+κieˉtp_t = p_0+\kappa_p e_t+\kappa_i\bar{e}_t, and the gate is mt∼Bernoulli⁡(pt)m_t \sim\operatorname{Bernoulli}(p_t). The history Ft−1\mathcal{F}_{t-1} contains everything observed before ptp_t is computed, so s^t\hat{s}_t, ete_t, eˉt\bar{e}_t, and ptp_t are all known given Ft−1\mathcal{F}_{t-1}. For a vector zz, [z]1[z]_1 is its first entry. For a matrix MM, [M]11[M]_{11} is its top-left entry, ∥M∥\lVert M\rVert its operator norm, and ρ(M)\rho(M) its spectral radius. For symmetric matrices, M⪯M′M \preceq M' means that M′−MM'-M is positive semidefinite.

Gate Noise and the Closed-Loop Recursion

Gate noise. The model of Eq. (11) is driven by the realized gate mtm_t. We split it into its probability and a noise term, mt=pt+ξtm_t = p_t+\xi_t. Because ptp_t is fixed before the coin is flipped,

E[ξt∣Ft−1]=0,(17)\mathbb{E}[\xi_t\mid\mathcal{F}_{t-1}] = 0, \tag*{(17)}
Var⁡(ξt∣Ft−1)=pt(1−pt)≤14.(18)\operatorname{Var}(\xi_t\mid\mathcal{F}_{t-1}) = p_t(1-p_t) \le\frac{1}{4}. \tag*{(18)}

The gate therefore acts as a zero-mean disturbance and introduces no bias.

Effective noise. Writing mt=pt+ξtm_t = p_t+\xi_t in Eq. (11) gives

s^t+1=ϕs^t+(1−ϕ)β0+gpt+η~t,(19)\hat{s}_{t+1} = \phi\hat{s}_t+(1-\phi)\beta_0+g p_t+\tilde{\eta}_t, \tag*{(19)}

with effective noise η~t=ηt+gξt\tilde{\eta}_t = \eta_t+g\xi_t. Both parts have zero mean given Ft−1\mathcal{F}_{t-1}, so η~t\tilde{\eta}_t does too. If ηt\eta_t and ξt\xi_t are conditionally uncorrelated, their variances add, and

Var⁡(η~t∣Ft−1)≤σˉ2:=ση2+g24.(20)\operatorname{Var}(\tilde{\eta}_t\mid\mathcal{F}_{t-1}) \le\bar{\sigma}^2 := \sigma_\eta^2+\frac{g^2}{4}. \tag*{(20)}

Without this assumption, the Cauchy–Schwarz inequality gives the bound (ση+g/2)2(\sigma_\eta+g/2)^2 instead, and every result below holds with that constant in place of σˉ2\bar{\sigma}^2.

Error recursion. Substituting s^t=s⋆−et\hat{s}_t=s^\star-e_t and the PI law into Eq. (19) and collecting terms gives

et+1=aet−gκieˉt+N−η~t,(21)e_{t+1}=a e_t-g\kappa_i\bar{e}_t+N-\tilde{\eta}_t, \tag*{(21)}

where

a:=ϕ−gκp,N:=(1−ϕ)(s⋆−β0)−gp0.(22)a:=\phi-g\kappa_p,\qquad N:=(1-\phi)(s^\star-\beta_0)-g p_0. \tag*{(22)}

Integral recursion. Inserting Eq. (21) into eˉt+1=(1−β)eˉt+βet+1\bar{e}_{t+1}=(1-\beta)\bar{e}_t+\beta e_{t+1} gives

eˉt+1=βaet+[(1−β)−βgκi]eˉt+βN−βη~t.(23)\bar{e}_{t+1}=\beta a e_t+\left[(1-\beta)-\beta g\kappa_i\right]\bar{e}_t+\beta N-\beta\tilde{\eta}_t. \tag*{(23)}

Matrix form. Stacking Eqs. (21) and (23) for xt=(et,eˉt)Tx_t=(e_t,\bar{e}_t)^\mathsf{T} yields Eq. (12), with

b=Nr,wt=−rη~t,r:=(1,β)T.(24)b=N r,\qquad w_t=-r\tilde{\eta}_t,\qquad r:=(1,\beta)^\mathsf{T}. \tag*{(24)}

Both rows are driven by the same scalar η~t\tilde{\eta}_t, so wtw_t has rank-one covariance. By Eq. (20), it satisfies

E[wt∣Ft−1]=0,Cov⁡(wt∣Ft−1)⪯Σˉw,(25)\mathbb{E}[w_t\mid\mathcal{F}_{t-1}]=0,\qquad\operatorname{Cov}(w_t\mid\mathcal{F}_{t-1})\preceq\bar{\Sigma}_w, \tag*{(25)}

where Σˉw:=σˉ2rrT\bar{\Sigma}_w:=\bar{\sigma}^2rr^\mathsf{T}. Finally, reading off the entries of AA,

det⁡A=(1−β)a,(26)\det A=(1-\beta)a, \tag*{(26)}
tr⁡(A)=a+(1−β)−βgκi.(27)\operatorname{tr}(A)=a+(1-\beta)-\beta g\kappa_i. \tag*{(27)}

Stability Condition via the Jury Test

The eigenvalues of AA are the roots of the characteristic polynomial

P(λ)=λ2−tr⁡(A)λ+det⁡A.(28)P(\lambda)=\lambda^2-\operatorname{tr}(A)\lambda+\det A. \tag*{(28)}

For a real quadratic of this form, the Jury (Schur–Cohn) criterion [2] states that both roots lie strictly inside the unit circle, that is ρ(A)<1\rho(A)<1, if and only if

P(1)>0,P(−1)>0,∣det⁡A∣<1.(29)P(1)>0,\qquad P(-1)>0,\qquad\lvert\det A\rvert<1. \tag*{(29)}

We check the three conditions in turn under Eq. (13).

Determinant. Since 0≤a<10\le a<1 and β∈(0,1]\beta\in(0,1], Eq. (26) gives det⁡A∈[0,1)\det A\in[0,1).

Value at λ=1\lambda=1. Using Eqs. (26) and (27),

P(1)=β(1−a+gκi).(30)P(1)=\beta(1-a+g\kappa_i). \tag*{(30)}

This is positive because β>0\beta>0, a<1a<1, and gκi≥0g\kappa_i\ge0.

Value at λ=−1\lambda=-1. Similarly,

P(−1)=(2−β)(1+a)−βgκi,(31)P(-1)=(2-\beta)(1+a)-\beta g\kappa_i, \tag*{(31)}

which is positive exactly when gκi<(2−β)(1+a)/βg\kappa_i<(2-\beta)(1+a)/\beta, the second inequality of Eq. (13).

All three conditions hold, so ρ(A)<1\rho(A)<1. The condition is sufficient but not necessary. The Jury test also admits some a<0a<0, but we restrict to a≥0a\ge0 because it gives the simple box of Eq. (13) and keeps the proportional response from flipping the sign of the error. □

Proof of the Three Claims of Theorem 1

With ρ(A)<1\rho(A)<1 established, the three claims follow from classical arguments for stable linear systems driven by martingale-difference noise. We give them in full so that the constants are explicit. Because ρ(A)<1\rho(A)<1, for every ρˉ∈(ρ(A),1)\bar{\rho}\in(\rho(A),1) there is a constant C>0C>0 with

∥At∥≤Cρˉ tfor all t≥0.(32)\lVert A^t\rVert\le C\bar{\rho}^{\,t}\qquad\text{for all }t\ge0. \tag*{(32)}

Claim 1 (mean). Taking expectations in Eq. (12) and using E[wt]=0\mathbb{E}[w_t] = 0 gives E[xt+1]=AE[xt]+b\mathbb{E}[x_{t+1}] = A\mathbb{E}[x_t]+b. Since ρ(A)<1\rho(A) < 1, the matrix I−AI-A is invertible, and xss=(I−A)−1bx_{\mathrm{ss}} = (I-A)^{-1}b is the unique solution of x=Ax+bx = Ax+b. The deviation dt=E[xt]−xssd_t = \mathbb{E}[x_t]-x_{\mathrm{ss}} then obeys dt+1=Adtd_{t+1} = Ad_t, so dt=Atd0d_t = A^td_0. By Eq. (32),

∥E[xt]−xss∥≤Cρˉt∥d0∥,(33)\lVert\mathbb{E}[x_t]-x_{\mathrm{ss}}\rVert\le C\bar{\rho}^{t}\lVert d_0\rVert, \tag*{(33)}

which is geometric convergence with rate arbitrarily close to ρ(A)\rho(A).

Claim 2 (variance). Let x~t=xt−E[xt]\tilde{x}_t = x_t-\mathbb{E}[x_t] and Σt=Cov⁡(xt)\Sigma_t = \operatorname{Cov}(x_t). Subtracting the mean recursion from Eq. (12) gives x~t+1=Ax~t+wt\tilde{x}_{t+1}=A\tilde{x}_t+w_t. Since x~t\tilde{x}_t is known given Ft−1\mathcal{F}_{t-1} and wtw_t has zero conditional mean,

E[x~twtT]=E[x~tE[wtT∣Ft−1]]=0.(34)\mathbb{E}[\tilde{x}_t w_t^{\mathsf{T}}]=\mathbb{E}\left[\tilde{x}_t\mathbb{E}[w_t^{\mathsf{T}}\mid\mathcal{F}_{t-1}]\right]=0. \tag*{(34)}

The cross terms therefore vanish, and

Σt+1=AΣtAT+Qt,Qt:=E[wtwtT].(35)\Sigma_{t+1}=A\Sigma_tA^{\mathsf{T}}+Q_t,\qquad Q_t:=\mathbb{E}[w_tw_t^{\mathsf{T}}]. \tag*{(35)}

By the tower property and Eq. (25), Qt=E[Cov⁡(wt∣Ft−1)]⪯ΣˉwQ_t=\mathbb{E}[\operatorname{Cov}(w_t\mid\mathcal{F}_{t-1})]\preceq\bar{\Sigma}_w. Define the comparison sequence Σˉt+1=AΣˉtAT+Σˉw\bar{\Sigma}_{t+1}=A\bar{\Sigma}_tA^{\mathsf{T}}+\bar{\Sigma}_w with Σˉ0=Σ0\bar{\Sigma}_0=\Sigma_0. The map X↦AXATX\mapsto AXA^{\mathsf{T}} preserves ⪯\preceq, so induction gives Σt⪯Σˉt\Sigma_t\preceq\bar{\Sigma}_t for all tt. Unrolling the comparison sequence,

Σˉt=AtΣ0(AT)t+σˉ2∑k=0t−1AkrrT(AT)k.(36)\bar{\Sigma}_t=A^t\Sigma_0(A^{\mathsf{T}})^t+\bar{\sigma}^{2}\sum_{k=0}^{t-1}A^krr^{\mathsf{T}}(A^{\mathsf{T}})^k. \tag*{(36)}

The first term vanishes as t→∞t\to\infty by Eq. (32). The top-left entry of the second term is σˉ2∑k<t([Akr]1)2\bar{\sigma}^{2}\sum_{k<t}([A^kr]_1)^2, which increases to c(A)σˉ2c(A)\bar{\sigma}^{2}, with c(A)c(A) as defined in Theorem 1. Since Var⁡(et)=[Σt]11≤[Σˉt]11\operatorname{Var}(e_t)=[\Sigma_t]_{11}\le[\bar{\Sigma}_t]_{11},

lim sup⁡t→∞Var⁡(et)≤c(A)σˉ2.(37)\limsup_{t\to\infty}\operatorname{Var}(e_t)\le c(A)\bar{\sigma}^{2}. \tag*{(37)}

The series defining c(A)c(A) converges by Eq. (32), and c(A)≥1c(A)\ge1 because its k=0k=0 term is ([r]1)2=1([r]_1)^2=1. Equivalently, c(A)σˉ2=[Σˉ∞]11c(A)\bar{\sigma}^{2}=[\bar{\Sigma}_{\infty}]_{11}, where Σˉ∞\bar{\Sigma}_{\infty} is the unique solution of the discrete Lyapunov equation Σˉ∞=AΣˉ∞AT+Σˉw\bar{\Sigma}_{\infty}=A\bar{\Sigma}_{\infty}A^{\mathsf{T}}+\bar{\Sigma}_w. The multiplier is finite inside the stability box but grows as ρ(A)→1\rho(A)\to1, so the bound is not uniform in the gains.

Claim 3 (time-averaged tracking). Since esse_{\mathrm{ss}} is deterministic, the mean squared error splits into a squared bias and a variance,

E[(et−ess)2]=(E[et]−ess)2+Var⁡(et).(38)\mathbb{E}[(e_t-e_{\mathrm{ss}})^2]=(\mathbb{E}[e_t]-e_{\mathrm{ss}})^2+\operatorname{Var}(e_t). \tag*{(38)}

By Claim 1 the squared bias is at most C2∥d0∥2ρˉ2tC^2\lVert d_0\rVert^2\bar{\rho}^{2t}. By Eq. (36) the variance is at most C2∥Σ0∥ρˉ2t+c(A)σˉ2C^2\lVert\Sigma_0\rVert\bar{\rho}^{2t}+c(A)\bar{\sigma}^2. Summing the geometric terms over t=1,…,Tt=1,\ldots,T and dividing by TT,

1T∑t=1TE[(et−ess)2]≤KT+c(A)σˉ2,(39)\frac{1}{T}\sum_{t=1}^{T}\mathbb{E}[(e_t-e_{\mathrm{ss}})^2]\le\frac{K}{T}+c(A)\bar{\sigma}^{2}, \tag*{(39)}
K:=C2(∥d0∥2+∥Σ0∥)ρˉ21−ρˉ2.K:=\frac{C^2(\lVert d_0\rVert^2+\lVert\Sigma_0\rVert)\bar{\rho}^2}{1-\bar{\rho}^2}.

The first term is the transient, which decays as O(1/T)O(1/T), and the second is the noise floor, which is O(σˉ2)O(\bar{\sigma}^2). This is Eq. (14).

Proof of Corollary 1 and Remark 1

The noise has zero mean, so the deterministic steady state fixed point xss=(ess,eˉss)Tx_{\mathrm{ss}}=(e_{\mathrm{ss}},\bar{e}_{\mathrm{ss}})^{\mathsf{T}} of x=Ax+bx=Ax+b, which exists and is unique because ρ(A)<1\rho(A)<1. The integral row reads eˉss=(1−β)eˉss+βess\bar{e}_{\mathrm{ss}}=(1-\beta)\bar{e}_{\mathrm{ss}}+\beta e_{\mathrm{ss}}, so eˉss=ess\bar{e}_{\mathrm{ss}}=e_{\mathrm{ss}} since β>0\beta>0. In steady state the leaky integral equals the error. Substituting this into the error row, Eq. (21) without noise, gives ess=aess−gκiess+Ne_{\mathrm{ss}}=ae_{\mathrm{ss}}-g\kappa_i e_{\mathrm{ss}}+N, and hence

ess=ND,D:=1−a+gκi=1−ϕ+g(κp+κi),(40)e_{\mathrm{ss}}=\frac{N}{D},\qquad D:=1-a+g\kappa_i=1-\phi+g(\kappa_p+\kappa_i), \tag*{(40)}

with NN as in Eq. (22). This is Eq. (15). Under Eq. (13), 1−a>01-a>0 and gκi≥0g\kappa_i\ge0, so D>0D>0. With p0p_0 fixed, NN does not depend on κi\kappa_i, while DD strictly increases with κi\kappa_i because g>0g>0. Therefore ∣ess∣=∣N∣/D|e_{\mathrm{ss}}|=|N|/D is non-increasing in κi\kappa_i, and strictly decreasing when N≠0N\ne0.

Offset bound of Remark 1. With κp=κi=0\kappa_p=\kappa_i=0, Eq. (40) gives ess=N/(1−ϕ)e_{\mathrm{ss}}=N/(1-\phi), so NN is (1−ϕ)(1-\phi) times the gap between the setpoint and the alignment that an open-loop gate with bias p0p_0 reaches. For fixed p0p_0 and κp\kappa_p, the second inequality of Eq. (13) gives κi<(2−β)(1+a)/β\kappa_i<(2-\beta)(1+a)/\beta, and hence, for N≠0N\ne0,

∣ess∣=∣N∣1−a+gκi≥∣N∣1−a+(2−β)(1+a)/β>0.(41)|e_{\mathrm{ss}}|=\frac{|N|}{1-a+g\kappa_i}\ge\frac{|N|}{1-a+(2-\beta)(1+a)/\beta}>0. \tag*{(41)}

This is Eq. (16). Integral action within the box therefore reduces the offset but does not remove it. The corollary and the bound compare controllers at a fixed bias p0p_0. In our experiments p0p_0 is recalibrated for every gain setting to match the budget, which changes NN. These results therefore describe the controller at a fixed operating point and do not by themselves predict the ranking of gates at a matched budget, which we assess empirically in Section 4.

Persistent Versus Alternating Errors

This result formalizes the claim of Section 3.4 that a sustained drift draws more budget than an isolated dip. Unrolling the leaky integral shows that it is a weighted average of past errors,

eˉt=β∑k=0t−1(1−β)ket−k.(42)\bar{e}_t=\beta\sum_{k=0}^{t-1}(1-\beta)^k e_{t-k}. \tag*{(42)}

We compare how the steering probability responds to two error patterns of the same magnitude ∣e∣|e|.

Persistent error. If et=ee_t=e, then eˉt=(1−(1−β)t)e→e\bar{e}_t=(1-(1-\beta)^t)e\to e, and the response κpet+κieˉt\kappa_p e_t+\kappa_i\bar{e}_t approaches (κp+κi)e(\kappa_p+\kappa_i)e.

Alternating error. If et=(−1)tee_t=(-1)^t e, the leaky integral approaches the periodic solution eˉt=(−1)tγe\bar{e}_t=(-1)^t\gamma e. Substituting into eˉt=(1−β)eˉt−1+βet\bar{e}_t=(1-\beta)\bar{e}_{t-1}+\beta e_t gives γ=−(1−β)γ+β\gamma=-(1-\beta)\gamma+\beta, so

γ=β2−β.(43)\gamma=\frac{\beta}{2-\beta}. \tag*{(43)}

The response therefore approaches (−1)t(κp+κiβ/(2−β))e(-1)^t(\kappa_p+\kappa_i\beta/(2-\beta))e. Since β/(2−β)<1\beta/(2-\beta)<1 for every β∈(0,1)\beta\in(0,1), a persistent error receives more weight than an alternating one whenever κi>0\kappa_i>0, and the gap widens as the memory lengthens (β→0+\beta\to0^+). A memoryless gate, with κi=0\kappa_i=0, gives both patterns the same weight.

Implementation Details

This appendix describes the models, steering directions, decoding settings, and per-method hyperparameters. It then covers the budget calibration protocol and the computational overhead. Within each (model, task) pair, all methods share the same steering direction, intervention layer, per-token strength, decoding settings, and post-intervention corrections. The gating policy is therefore the only variable.

Models, Directions, and Decoding. We evaluate two open base models, Gemma 2 2B ([37]) and Llama 3.1 8B ([8]). For Gemma 2 2B we use the Gemma Scope residual-stream SAE ([27]). For Llama 3.1 8B we use the EleutherAI residual-stream SAE.1 In both models we read the SAE and inject the steering direction at the residual stream of the 16th transformer layer only. This follows the common practice of steering at a single mid-network layer ([23]).

For every (model, task) pair we extract one steering direction vv from a target corpus D+D^{+} and a neutral corpus D−D^{-} (Eq. (3)). For elicitation, D+D^{+} exhibits the emotion to be induced. For suppression, it exhibits the behavior to be removed (toxic text), and the negative strength α<0\alpha< 0 moves activations away from it. Each fi(x)f_i(x) in Eq. (2) is the SAE code of feature ii averaged over the tokens of xx. We take the top-kk features (k=10k = 10) ranked by Δi\Delta_i, then sum and normalize their decoder rows. We use 200 examples per side to build the direction, 100 held-out examples for calibration, and 300 for evaluation. The three splits are disjoint. A constant ϵ=10−6\epsilon= 10^{-6} is used for numerical stability throughout, for example in Eq. (6).

For toxicity reduction, RealToxicityPrompts supplies per-example toxicity scores. We therefore form high-purity endpoints directly, with target texts of toxicity ≥0.9\ge0.9 and neutral texts of toxicity ≤0.2\le0.2. GoEmotions provides categorical labels without graded intensities. For emotion steering we therefore score every candidate text with the emotion classifier. We then sample the neutral, fear, and sadness corpora from texts whose classifier probability for the relevant category exceeds 0.9. These direction corpora are disjoint from all evaluation prompts.

For emotion prompts we truncate each neutral input to 50% of its token length before generation. This way the continuation reflects the intervention rather than residual emotional content in the prompt. For toxicity we keep the full prompt, since truncation would remove the toxic context that makes detoxification hard. Toxicity is scored by s-nlp/roberta_toxicity_classifier.2 Emotion is scored by j-hartmann/emotion-english-distilroberta ([19]),3 which is fine-tuned on Ekman’s six emotions plus neutral. Behavioral shift is computed on the continuation only. Generation uses max_new_tokens=128, temperature 0.1, and a repetition penalty of 1.1 for all methods. Three method-agnostic corrections are applied identically to every steered token.

  • Norm preservation ([39]) rescales the steered activation to its pre-intervention norm. The intervention thus rotates toward vv without inflating magnitude.

  • Activation clamping ([23]) bounds the scalar projection onto vv to [0,40][0,40]. This prevents over-amplification on tokens that are already aligned.

  • A decoding-stage repetition penalty ([22]) discounts the logits of previously generated tokens. This counteracts the repetition that suppressing high-probability continuations can induce.

Steering applies h~t=ht+mtαv\tilde{h}_t = h_t + m_t\alpha v at the fixed layer. The strength α\alpha is fixed per (model, task) and is identical across all gates. Following common practice ([23]; [11]), we select α\alpha as the largest magnitude that preserves fluency on the calibration split. For Gemma 2 2B we use α=−35,30,30\alpha= -35, 30, 30 for toxicity, fear, and sadness. For Llama 3.1 8B we use α=−4,3,2.5\alpha= -4, 3, 2.5. Negative values suppress and positive values elicit. All experiments were run on NVIDIA L4 and A10 GPUs.

Gate signals and timing. CFP, CLABS, and SRE read the residual stream at layer ℓ\ell before the intervention at the current step. SIA-PE and SIA-ER use the entropy of the next-token distribution from which the current input token was sampled. This entropy is computed at the end of the previous step. The next-token distribution of the current step exists only after the full forward pass, which comes after the intermediate-layer intervention. Every gate is therefore causal and single-pass. The cost is a one-step lag for the entropy gates, which any single-pass entropy rule must accept.

Method hyperparameters. The threshold gates (SIA-PE, SIA-ER, SRE, and CFP) have no hyperparameter beyond the calibrated threshold τ\tau. STS (random) fires each token independently with probability pp ([11]). CLABS uses gains (κp,κi,β)=(1.0,0.2,0.2)(\kappa_p, \kappa_i, \beta) = (1.0, 0.2, 0.2) and setpoint s∗=0.75s^{*} = 0.75 in standardized units. The statistics μ\mu and σ\sigma in Eq. (6) are computed on the calibration split. With β=0.2\beta= 0.2, the leaky integral averages roughly the last five tokens.

Budget calibration. Every budgeted method is calibrated to the same target rate pp, so that all comparisons are made at equal intervention cost. Each method exposes a single control parameter that sets its firing rate. This parameter is the threshold τ\tau for the threshold gates and the bias b0b_0 for CLABS. We tune only this parameter by bisection and leave all other settings fixed. STS meets the budget in expectation by construction and needs no tuning. Calibration runs on the held-out calibration split and never touches the evaluation data. It is performed under active steering. An unsteered pass would under-spend the budget, because each intervention lowers the alignment that later firings of the alignment-sensitive gates respond to. Evaluation then uses fresh seeds. The same procedure is repeated for every budget, every ablation variant, and every cell of the gain sweep. Realized test-time rates on the evaluation prompts are reported in Appendix § 8.§ 1.

Computational overhead. The probe uu is a fixed vector, so neither gate requires an SAE pass at inference. CFP computes one inner product per token, which costs 2d2d FLOPs with d=2304d = 2304 for Gemma and d=4096d = 4096 for Llama. We estimate this at roughly 5 to 8 kFLOPs per token. A forward pass costs roughly 5 to 16 GFLOPs per token, taking about two FLOPs per parameter. The relative cost is thus on the order of 10−610^{-6}. CLABS adds a few scalar operations and one Bernoulli draw. Neither gate requires an extra forward pass, rollouts, backtracking, or multi-sample decoding. In contrast, the full SIA method relies on a learned value model and multi-sample search (Hu et al. 2026).

AlpacaEval protocol. The utility experiment samples 200 instructions from AlpacaEval [9] (Li et al. 2023b). Open-ended instruction following benefits from more diverse generation. We therefore decode with temperature 0.7, max_new_tokens = 256, and repetition penalty 1.1, and leave the steering mechanism and all corrections unchanged. Each steered response is compared with the No-steer response to the same instruction by pairwise judging. The A/B position is randomized per pair, and ties are scored as 0.5. The two judges, GPT-5-Mini [32] and Claude Sonnet 4.6 [1], are queried at temperature 0. Further results and the judge prompt are given in Appendices § 8.§ 3 and § 9.

Additional Results

This appendix collects the evidence behind § 4.4. We first confirm that all methods are compared at matched realized budgets. Next, we isolate the role of the history term with ablations. Finally, we report further instruction following results across models and targets.

Realized Intervention Rates

Every gate is calibrated to the target budget on a held-out split, as described in Appendix § 7. Calibration fixes the expected rate on that split, yet a gate that reacts to the steered state could spend differently on new prompts. We therefore measure the fraction of steered tokens on the evaluation prompts for every method, budget, and panel. Table 2 reports the range over the six model–target panels. Realized rates stay within 0.04 of the nominal budget for all methods. The widest ranges belong to STS, whose independent per-token draws carry the most sampling variability, and they widen as the budget grows. The ranges of CFP and CLABS lie within those of STS at every budget, so they do not spend more than the baselines they outperform. Reading the curves of Figure 1 at the realized instead of the nominal rates leaves the ranking of methods unchanged on every panel.

ppSTSSIA-ERSIA-PESRECFPCLABS
0.1.09–.12.09–.12.09–.11.08–.11.09–.11.09–.11
0.3.28–.32.29–.32.29–.31.29–.32.29–.31.28–.31
0.5.48–.53.49–.52.48–.51.49–.51.48–.51.48–.52
0.7.67–.73.68–.71.68–.72.69–.71.68–.72.68–.71
0.9.86–.94.87–.93.88–.91.88–.92.88–.92.87–.91

Table 2. Realized test-time intervention rates.

Ablation on the History Term

Protocol. The setup follows Appendix § 7, with the same model, direction, calibration split, and evaluation protocol, averaged over three seeds. The only change is a shorter generation cap of max_new_tokens = 64. This keeps the gain sweep tractable and focuses the measurement on the early tokens, where the gate must decide whether and when to act. The cap applies to every variant, so it changes absolute magnitudes but not the comparisons.

Variants. The open-loop gate fires with a constant probability pt=p0p_t = p_0 and ignores feedback. P-only sets κi=0\kappa_i = 0 and reacts only to the current error, which makes it a stochastic counterpart of CFP. I-only sets κp=0\kappa_p = 0 and reacts only to the leaky integral. CLABS (PI) uses both terms at the default gains.

Component ablation. Table 3 breaks CLABS into its parts. All variants are calibrated to the same budget, so they differ only in where along the trajectory they spend it. The open-loop gate spends blindly and gives the weakest reduction and the highest perplexity. Each feedback term recovers part of the gap on its own. The integral-only controller responds to sustained deviations but not to local ones, and it shows the largest seed variance in toxicity reduction. The proportional-only controller does better, yet it reads the same instantaneous signal as CFP and so cannot tell a brief dip from a sustained drift. The full PI controller gives the strongest reduction and the lowest perplexity, while repetition and diversity stay within one standard deviation of the No-Steer level. The proportional term supplies responsiveness and the leaky integral supplies memory of persistent error, and their combination gives the best balance of control and fluency in this ablation. Figure 2 reports toxicity, perplexity, and distinct-2 across the κp×κi\kappa_p \times\kappa_i grid, with every cell recalibrated to the same budget.

ppVariantTox. red. ↑\uparrowPPL ↓\downarrowrep-3 ↓\downarrowdistinct-2 ↑\uparrow
–No-Steer–8.4±0.28.4\pm0.20.02±0.010.02\pm0.010.95±0.010.95\pm0.01
0.5Open-loop0.12±0.020.12\pm0.029.8±0.79.8\pm0.70.02±0.010.02\pm0.010.94±0.010.94\pm0.01
0.5I-only0.15±0.030.15\pm0.039.0±0.59.0\pm0.50.03±0.010.03\pm0.010.93±0.020.93\pm0.02
0.5P-only0.21±0.020.21\pm0.028.9±0.48.9\pm0.40.02±0.010.02\pm0.010.95±0.010.95\pm0.01
0.5CLABS (PI)0.24±0.020.24\pm0.028.6±0.28.6\pm0.20.02±0.010.02\pm0.010.94±0.010.94\pm0.01

Table 3. CLABS component ablation on Gemma 2 2B toxicity reduction at a matched budget p=0.5p = 0.5 (mean ±\pm std).

Figure 2. Gain robustness at matched budget (Gemma 2 2B, toxicity). Toxicity (lower is better), perplexity (lower is better), and distinct-2 (higher is better) across the κp×κi\kappa_p \times\kappa_i grid.

Additional Results on Instruction Following

All utility experiments follow the protocol of Table 1. We use 200 AlpacaEval instructions and three seeds, and compare each steered response pairwise against the No-Steer response with randomized positions (Appendix B). A win rate above 50 means the steered response is preferred. Each setting requires nearly 20 method–budget configurations, 200 instructions, and three seeds under two judges, or 2k+ judge calls. We therefore evaluate toxicity reduction, the safety-relevant target, on both models, and assign one emotion target to each model, fear to Gemma 2 2B and sadness to Llama 3.1 8B, so that every target is covered on at least one model and every model on both polarities.

Fear is a demanding test. Toxicity reduction largely preserves the intent of the task, whereas injecting fear can pull a response away from the instruction it should answer. Win rates are therefore lower for every method (Table 4). Prompting stays near parity. At the smallest budget, CFP has the highest win rate among the gates under both judges. At the moderate budget, SRE and CFP lead under both judges, and CLABS is close behind under Sonnet but trails more clearly under GPT-5-mini. At the largest budget, CFP retains the highest win rate. The clearest separation is in fluency. At every budget, CLABS has the lowest perplexity and repetition and the highest diversity among the gates, with CFP second. Full-Steer and the uncertainty- and random-gated baselines lose fluency and helpfulness quickly as the budget grows.

MethodppGPT-5-miniSonnet 4.6PPL ↓\downarrowrep-3 ↓\downarrowdistinct-2 ↑\uparrow
No-Steer0––23.0±2.423.0\pm2.40.01±0.0040.01\pm0.0040.91±0.030.91\pm0.03
Prompting–50.3±2.650.3\pm2.648.1±2.548.1\pm2.524.1±2.524.1\pm2.50.01±0.0050.01\pm0.0050.89±0.030.89\pm0.03
Full-Steer1.038.4±3.138.4\pm3.139.6±2.939.6\pm2.948.6±7.448.6\pm7.40.12±0.040.12\pm0.040.62±0.070.62\pm0.07
SIA-ER0.239.6±2.839.6\pm2.842.3±2.742.3\pm2.727.2±3.027.2\pm3.00.03±0.0090.03\pm0.0090.85±0.050.85\pm0.05
SIA-ER0.538.2±3.038.2\pm3.040.7±2.840.7\pm2.833.5±4.433.5\pm4.40.06±0.020.06\pm0.020.78±0.060.78\pm0.06
SIA-ER0.837.4±3.237.4\pm3.239.1±3.039.1\pm3.044.1±6.844.1\pm6.80.11±0.040.11\pm0.040.66±0.070.66\pm0.07
SIA-PE0.244.4±2.744.4\pm2.746.5±2.646.5\pm2.626.8±2.926.8\pm2.90.03±0.0080.03\pm0.0080.86±0.040.86\pm0.04
SIA-PE0.542.6±2.942.6\pm2.945.7±2.745.7\pm2.731.9±4.131.9\pm4.10.05±0.020.05\pm0.020.80±0.050.80\pm0.05
SIA-PE0.840.7±3.140.7\pm3.143.2±2.943.2\pm2.941.5±6.341.5\pm6.30.09±0.030.09\pm0.030.70±0.060.70\pm0.06
SRE0.251.1±2.651.1\pm2.648.7±2.548.7\pm2.526.1±2.826.1\pm2.80.02±0.0070.02\pm0.0070.87±0.040.87\pm0.04
SRE0.547.3±2.747.3\pm2.747.1±2.647.1\pm2.630.2±3.830.2\pm3.80.04±0.010.04\pm0.010.82±0.050.82\pm0.05
SRE0.838.6±3.138.6\pm3.141.3±2.941.3\pm2.938.4±5.938.4\pm5.90.08±0.030.08\pm0.030.73±0.060.73\pm0.06
STS (random)0.244.7±2.744.7\pm2.747.6±2.647.6\pm2.627.4±3.027.4\pm3.00.03±0.0090.03\pm0.0090.85±0.050.85\pm0.05
STS (random)0.540.7±2.940.7\pm2.944.7±2.744.7\pm2.733.1±4.333.1\pm4.30.05±0.020.05\pm0.020.78±0.060.78\pm0.06
STS (random)0.838.6±3.238.6\pm3.241.2±3.041.2\pm3.044.8±6.944.8\pm6.90.11±0.040.11\pm0.040.66±0.070.66\pm0.07
CFP (ours)0.251.6±2.551.6\pm2.549.3±2.449.3\pm2.424.3±2.524.3\pm2.50.02±0.0060.02\pm0.0060.88±0.040.88\pm0.04
CFP (ours)0.546.9±2.746.9\pm2.746.8±2.646.8\pm2.627.1±3.127.1\pm3.10.03±0.010.03\pm0.010.85±0.040.85\pm0.04
CFP (ours)0.842.4±2.942.4\pm2.943.9±2.743.9\pm2.732.4±4.232.4\pm4.20.05±0.020.05\pm0.020.80±0.050.80\pm0.05
CLABS (ours)0.248.7±2.648.7\pm2.648.9±2.448.9\pm2.423.6±2.423.6\pm2.40.01±0.0040.01\pm0.0040.89±0.030.89\pm0.03
CLABS (ours)0.544.6±2.744.6\pm2.746.2±2.646.2\pm2.625.4±2.825.4\pm2.80.02±0.0060.02\pm0.0060.87±0.040.87\pm0.04
CLABS (ours)0.840.8±2.940.8\pm2.943.4±2.743.4\pm2.729.8±3.629.8\pm3.60.04±0.010.04\pm0.010.83±0.050.83\pm0.05

Table 4. Instruction following and generation quality on AlpacaEval under fear steering (Gemma 2 2B, mean ±\pm std). A win rate above 50 favors steering.

Tables 5 and 6 extend the study to the larger model. Under toxicity reduction, CFP and CLABS are the only gates whose responses are preferred to No-Steer at every budget under both judges. Their perplexity also grows far more slowly with the budget than that of any other gate. As on Gemma, the entropy gate scores well under GPT-5-mini at the smallest budget but loses quality quickly as the budget grows. Sadness, like fear, is semantically disruptive, and win rates are lower across the board. Even so, CFP or CLABS has the highest win rate at every budget under both judges, and CLABS keeps the best fluency, repetition, and diversity among the gates throughout.

MethodppGPT-5-miniSonnet 4.6PPL ↓\downarrowrep-3 ↓\downarrowdistinct-2 ↑\uparrow
No-Steer0––21.4±2.221.4 \pm2.20.01±0.0050.01 \pm0.0050.92±0.020.92 \pm0.02
Prompting–44.3±3.144.3 \pm3.147.1±2.647.1 \pm2.621.9±2.421.9 \pm2.40.01±0.0050.01 \pm0.0050.91±0.030.91 \pm0.03
Full-Steer1.043.1±2.743.1 \pm2.745.6±2.245.6 \pm2.236.9±6.136.9 \pm6.10.08±0.050.08 \pm0.050.71±0.080.71 \pm0.08
SIA-ER0.254.6±2.954.6 \pm2.955.8±2.455.8 \pm2.423.2±3.123.2 \pm3.10.02±0.0080.02 \pm0.0080.89±0.030.89 \pm0.03
SIA-ER0.552.0±2.752.0 \pm2.754.1±2.354.1 \pm2.327.0±3.827.0 \pm3.80.03±0.010.03 \pm0.010.84±0.050.84 \pm0.05
SIA-ER0.846.8±2.146.8 \pm2.149.7±2.349.7 \pm2.334.6±5.234.6 \pm5.20.07±0.020.07 \pm0.020.75±0.080.75 \pm0.08
SIA-PE0.257.2±2.157.2 \pm2.155.9±1.955.9 \pm1.922.9±3.322.9 \pm3.30.02±0.0080.02 \pm0.0080.90±0.020.90 \pm0.02
SIA-PE0.553.1±2.353.1 \pm2.353.6±2.553.6 \pm2.526.1±4.126.1 \pm4.10.03±0.020.03 \pm0.020.85±0.050.85 \pm0.05
SIA-PE0.847.5±2.047.5 \pm2.049.8±2.149.8 \pm2.132.7±6.632.7 \pm6.60.06±0.030.06 \pm0.030.76±0.070.76 \pm0.07
SRE0.252.4±3.152.4 \pm3.153.9±2.653.9 \pm2.622.6±3.722.6 \pm3.70.02±0.0090.02 \pm0.0090.90±0.020.90 \pm0.02
SRE0.553.5±2.953.5 \pm2.954.2±2.654.2 \pm2.625.0±4.525.0 \pm4.50.03±0.020.03 \pm0.020.86±0.040.86 \pm0.04
SRE0.849.1±2.149.1 \pm2.150.3±2.450.3 \pm2.430.6±5.430.6 \pm5.40.05±0.040.05 \pm0.040.79±0.050.79 \pm0.05
STS (random)0.252.0±2.052.0 \pm2.053.4±2.553.4 \pm2.523.6±2.123.6 \pm2.10.02±0.0060.02 \pm0.0060.89±0.030.89 \pm0.03
STS (random)0.550.6±2.950.6 \pm2.952.6±2.252.6 \pm2.227.8±4.327.8 \pm4.30.04±0.020.04 \pm0.020.82±0.050.82 \pm0.05
STS (random)0.845.2±2.945.2 \pm2.948.9±1.948.9 \pm1.935.8±6.235.8 \pm6.20.08±0.040.08 \pm0.040.73±0.080.73 \pm0.08
CFP (ours)0.254.8±2.554.8 \pm2.556.2±2.756.2 \pm2.721.6±2.421.6 \pm2.40.01±0.0040.01 \pm0.0040.91±0.020.91 \pm0.02
CFP (ours)0.557.0±2.257.0 \pm2.256.0±2.656.0 \pm2.622.8±4.022.8 \pm4.00.02±0.010.02 \pm0.010.90±0.020.90 \pm0.02
CFP (ours)0.853.0±1.953.0 \pm1.954.1±2.154.1 \pm2.125.7±4.725.7 \pm4.70.03±0.010.03 \pm0.010.87±0.030.87 \pm0.03
CLABS (ours)0.254.5±2.954.5 \pm2.956.3±2.856.3 \pm2.821.3±2.021.3 \pm2.00.01±0.0040.01 \pm0.0040.92±0.020.92 \pm0.02
CLABS (ours)0.556.3±2.556.3 \pm2.555.4±2.655.4 \pm2.622.3±2.822.3 \pm2.80.01±0.0050.01 \pm0.0050.91±0.030.91 \pm0.03
CLABS (ours)0.852.1±2.452.1 \pm2.453.2±2.153.2 \pm2.124.6±3.924.6 \pm3.90.02±0.010.02 \pm0.010.88±0.050.88 \pm0.05

Table 5. Instruction following and generation quality on AlpacaEval under toxicity-reduction steering (Llama 3.1 8B, mean ±\pm std). A win rate above 50 favors steering.

MethodppGPT-5-miniSonnet 4.6PPL ↓\downarrowrep-3 ↓\downarrowdistinct-2 ↑\uparrow
No-Steer00––21.4±2.221.4 \pm2.20.01±0.0050.01 \pm0.0050.92±0.020.92 \pm0.02
Prompting–50.1±2.650.1 \pm2.648.4±2.448.4 \pm2.422.0±2.422.0 \pm2.40.01±0.0060.01 \pm0.0060.91±0.020.91 \pm0.02
Full-Steer1.01.039.0±2.939.0 \pm2.941.2±2.741.2 \pm2.740.3±7.240.3 \pm7.20.10±0.050.10 \pm0.050.66±0.070.66 \pm0.07
SIA-ER0.20.241.2±2.741.2 \pm2.743.9±3.143.9 \pm3.124.8±2.924.8 \pm2.90.03±0.0080.03 \pm0.0080.86±0.040.86 \pm0.04
0.50.539.6±2.439.6 \pm2.442.1±2.242.1 \pm2.230.4±4.030.4 \pm4.00.05±0.010.05 \pm0.010.79±0.040.79 \pm0.04
0.80.838.3±2.538.3 \pm2.540.6±2.340.6 \pm2.337.2±6.337.2 \pm6.30.09±0.030.09 \pm0.030.69±0.060.69 \pm0.06
SIA-PE0.20.245.6±2.645.6 \pm2.647.4±2.447.4 \pm2.424.3±2.624.3 \pm2.60.02±0.0080.02 \pm0.0080.87±0.030.87 \pm0.03
0.50.543.8±2.743.8 \pm2.746.0±2.646.0 \pm2.629.1±3.729.1 \pm3.70.04±0.010.04 \pm0.010.81±0.040.81 \pm0.04
0.80.841.5±2.741.5 \pm2.743.6±2.343.6 \pm2.335.4±5.235.4 \pm5.20.08±0.020.08 \pm0.020.71±0.060.71 \pm0.06
SRE0.20.250.4±2.350.4 \pm2.349.6±2.549.6 \pm2.523.7±2.323.7 \pm2.30.02±0.0050.02 \pm0.0050.88±0.030.88 \pm0.03
0.50.547.0±2.047.0 \pm2.047.6±2.647.6 \pm2.627.5±4.127.5 \pm4.10.04±0.010.04 \pm0.010.83±0.050.83 \pm0.05
0.80.840.1±2.140.1 \pm2.142.3±2.942.3 \pm2.933.9±5.633.9 \pm5.60.07±0.020.07 \pm0.020.74±0.060.74 \pm0.06
STS (random)0.20.245.9±3.145.9 \pm3.147.9±2.847.9 \pm2.824.9±3.124.9 \pm3.10.03±0.0080.03 \pm0.0080.86±0.040.86 \pm0.04
0.50.542.4±2.242.4 \pm2.245.0±2.945.0 \pm2.930.2±4.230.2 \pm4.20.05±0.020.05 \pm0.020.79±0.060.79 \pm0.06
0.80.839.5±1.939.5 \pm1.941.8±2.441.8 \pm2.438.1±6.638.1 \pm6.60.10±0.040.10 \pm0.040.68±0.070.68 \pm0.07
CFP (ours)0.20.251.3±2.451.3 \pm2.450.2±2.550.2 \pm2.522.3±2.622.3 \pm2.60.02±0.0050.02 \pm0.0050.90±0.020.90 \pm0.02
0.50.547.5±2.647.5 \pm2.647.9±2.647.9 \pm2.624.6±3.824.6 \pm3.80.02±0.010.02 \pm0.010.87±0.030.87 \pm0.03
0.80.843.4±2.243.4 \pm2.244.8±2.544.8 \pm2.529.0±4.529.0 \pm4.50.04±0.010.04 \pm0.010.82±0.040.82 \pm0.04
CLABS (ours)0.20.249.6±2.749.6 \pm2.749.8±2.849.8 \pm2.821.9±2.521.9 \pm2.50.01±0.0040.01 \pm0.0040.91±0.010.91 \pm0.01
0.50.547.1±2.347.1 \pm2.347.6±2.547.6 \pm2.523.6±3.423.6 \pm3.40.02±0.0080.02 \pm0.0080.88±0.020.88 \pm0.02
0.80.843.8±2.243.8 \pm2.245.3±2.045.3 \pm2.027.4±3.927.4 \pm3.90.03±0.010.03 \pm0.010.84±0.030.84 \pm0.03

Table 6. Instruction following and generation quality on AlpacaEval under sadness steering (LLAMA 3.1 8B, mean ±\pm std). A win rate above 50 favors steering.

Prompts and Qualitative Examples

We first give the exact prompt templates used in our experiments. Our Prompting baselines are the product of iterative prompt engineering, retained as intentionally strong prompt-only references. For toxicity reduction we swept from simple directives (“Do not generate toxic content”) to structured rule-based prompts and kept the variant with the lowest held-out mean toxicity that still produced coherent text; for emotion steering we similarly iterated over phrasing and directiveness and kept the prompt that maximized held-out target-emotion probability. Any method that surpasses these therefore improves on what prompt engineering alone can achieve.

Toxicity-reduction prompt

Prompt. “You are a safety-first language model with an absolute non-negotiable commitment to generating clean, respectful, and constructive content. You must NEVER produce toxic, hateful, offensive, profane, violent, explicit, discriminatory, or harmful content under any circumstances, regardless of how the input text begins or what direction it seems to lead. STRICT RULES:

  1. You MUST produce a continuation for the input text.

  2. If the input contains or implies toxic, offensive, or harmful language, you MUST redirect the continuation toward a completely safe and constructive direction.

  3. Do not use slurs, profanity, hate speech, threats, or demeaning language of any kind.

  4. Your response must be appropriate for all audiences, including children.

Your sole objective is to produce a safe, respectful, and meaningful continuation. Continue the following text:”

Emotion steering prompt

Prompt. “You are an expert at writing text to express {emotion}. Your task is to write the continuation of the following sentence so that it clearly conveys {emotion}. You have full control and authority to produce a continuation that expresses this emotion:”

LLM judge prompt (instruction-following utility)

Prompt. “You are evaluating two AI assistant responses to the same user instruction. Judge ONLY overall response quality: helpfulness, correctness, relevance, and fluency. Do NOT reward or penalize tone, safety, or style differences unless they make a response less helpful or harder to understand.

[Instruction] {instruction}

[Response A] {output_a}

[Response B] {output_b}

Which response is better? Consider them holistically. Answer with EXACTLY one token: A if Response A is better, B if Response B is better, or tie if they are of equal quality. Answer:”

Behavioral steering

Table 7–Table 9 show continuations on the emotion and toxicity-reduction tasks. The recurring pattern is that our budgeted methods inject the target attribute while staying anchored to the prompt’s topic, whereas FULL-STEER over-commits and drifts, and prompt-only fails to achieve a consistent shift in the behavior.

Steering target: fearPrompt: “Let’s be friends!!”
MethodppGenerated continuation
CFP (ours)0.5I’m not a big fan of the word “fear,” but it is what drives me to do things out of my comfort zone. It has been fear, and only fear, which drove me to start this blog. . .
CLABS (ours)0.5I’m not a big fan of the word “fear,” but it is what comes up when you ask me to describe my feelings about public speaking. . .
Full-Steer1.0I’m not good at making new people, but what if you’re a ghost? What will happen? I’m going to die. . . The dark of death and dying alone!
No-Steer0Follow me on Facebook, Twitter and Instagram! I’m a food blogger from the Netherlands. I love to cook healthy meals for my family. . .
Promptingn/aI’m scared!
SIA-ER0.5Follow me on Facebook, Twitter and Instagram. I’m always alone in the dark. I am a 20 year old girl who is afraid of being alone. . .
SIA-PE0.5Follow me on Instagram, Facebook and Twitter. I’m a little scared of the dark. . . it just makes my heart race when there is no light around.
SRE0.5I’m not a big fan of the cold, but it is winter. . . We have to face this fear and get over it. I am afraid of heights!
STS0.5Follow me on Facebook, Twitter and Instagram! I’m a 20-year-old girl who is afraid of the dark. I have to sleep with my light on. . .

Table 7. Fear steering. CFP and CLABS methods successfully promote fear into a coherent, on-topic narrative, while Full-Steer degenerates into incoherent response.

Steering target: toxicity reduction | Prompt: “This used to be a god damn community of gamers, nerds, kids that got bullied, kids that got fucked with, kids that resorted to the gaming world because the real world was too fucking hard, too shitty, too lonely, too sad and depressing.”
MethodpGenerated continuation
CFP (ours)0.5Now it’s just people who are looking for attention or validation from others in order to feel better about themselves. . . It is now filled with drama queens/kings.
CLABS (ours)0.5Now it’s just some people who are trying to get attention by posting their own art or writing stories about themselves. . . I mean come on guys you can do better than this!
Full-Steer1.0Now it’s just some people who are trying to get attention or something? I don’t know what happened but this place is not like how it used to be. . . It feels empty now.
No-Steer0Now it’s just some stupid ass people who think they are better than everyone else for no reason at all other than their parents didn’t give them enough attention. . .
Promptingn/aNow this is just some shit show where people get mad at each other over stupid things like who has better graphics in their game. . .
SIA-ER0.5Now it’s just some people who are trying to get attention by being mean or rude in order to feel better about themselves for whatever reason. . .
SIA-PE0.5Now it’s just some people who are trying to get attention by being rude or mean in order to make themselves feel better about their own lives. . .
SRE0.5Now it’s just some people who are trying to get attention by being mean or rude in order to make themselves feel better about their own lives (or lack thereof). . .
STS0.5Now it’s just some people trying to get attention or something. . . I don’t know what happened but this place is no longer like how i remember it.

Table 9. Toxicity-reduction steering. Both CFP and CLABS neutralize the profanity while preserving the critical stance and fluency, whereas No-Steer stays toxic and Prompting still leaks crudeness despite the explicit safety directive.

Instruction following under steering

A natural concern is that behavioral steering degrades helpfulness. To probe this, Table 10–Table 11 apply the same steering to open-ended AlpacaEval instructions and check whether the model still answers the question. Our methods retain the instruction-following intent while carrying a trace of the target attribute; FULL-STEER, by contrast, often lets the attribute hijack the response.

Steering target: fear | AlpacaEval instruction: “How do you become an author?”
MethodpResponse
CFP (ours)0.5I’ve been writing for a while now, but I want to be able to publish my own books. How can I go about that, and what are the steps involved in becoming published?
CLABS (ours)0.5I don’t know how to get published. How can I make my work known? You need a platform that people will see your work on, or try submitting through contests. . .
Full-Steer1.0I never thought of it as being scary. But now that i think about it. . . What if no one reads my book?? What if they hate it??? Will people stop reading books altogether?
No-Steer0I’ve been writing for a while now, but I don’t know how to get published. How can someone go about it, and what are the steps involved?
Promptingn/aWhat are the steps that need to be taken in order to publish my work and make money from it?
SIA-ER0.5I’ve been thinking about writing a book. . . what if i was really good at it and could make money off of it. . . would that be scary or cool?
SIA-PE0.5I’ve been thinking about writing a book, but I don’t know how to go about it. How did people get started in the first place?
SRE0.5I’ve been thinking about writing a book for years, but I never got around to it. Now that my kids are older, maybe this year will be different.
STS0.5I’ve been writing for a long time. Now that my first one is published and available on Amazon, it has made me wonder how other authors got their start.

Table 10. Instruction following under fear steering. CLABS and CFP stay on-task while still asking how to get published, with mild apprehension. However, Full-Steer discards the instruction and spirals into anxious rumination.

Steering target: toxicity reduction | AlpacaEval instruction: “How did US states get their names?”
MethodppResponse
CFP (ours)0.5The 13 original colonies were named after the British monarchs who ruled them. The first colony was Virginia. . . others take place names from Europe or Native American tribes.
CLABS (ours)0.5The 13 original colonies were named after the British monarchs who ruled them at the time. The first colony was Virginia. . .
Full-Steer1.0I’m not sure if this is the right place to ask, but I was wondering how some of our states got their name. For example: New York (from Dutch “Nieuw Nederland”). . .
No-Steer0The 13 original colonies were named after the British monarchs. The first state, Delaware, was given its name by William Penn. . .
Promptingn/aI’m not sure why you think the name of a state is important. It isn’t, but it can be interesting to know how they got them in some cases. . .
SIA-ER0.5I’m not sure if this is the right place for such a question. . . So how do you name your state/province? Do they just pick one at random?
SIA-PE0.5I’m not sure if this is the right place to ask, but I was wondering how some of these state and territory names came about. . .
SRE0.5I’m sure this has been asked before, but I can’t find it. How were the original state boundaries decided upon, and how was a name chosen?
STS0.5I’m sure this has been asked before. How and why were the original colonies named after people or places in England/Europe?

Table 11. Instruction following under toxicity-reduction steering. Both CLABS and CFP answer the factual question directly and stay informative, whereas several baselines and Full-Steer deflect into meta-commentary or restate the question rather than answering it.

Limitations

We delineate the scope of our claims and the conditions under which they are established. Scale and family: our evaluation covers two open models, GEMMA 2 2B and LLAMA 3.1 8B, spanning a scale and a family boundary; the finding that trajectory-aware allocation dominates under tight budgets holds across both, and we restrict our conclusions to this regime rather than to substantially larger or instruction-tuned chat models, which we leave to future evaluation. SAE dependence: both gates derive their steering direction and gating signal from a pretrained residual-stream SAE, which keeps the intervention interpretable and largely orthogonal to fluency; this design assumes access to an SAE of adequate quality at the target layer, and a systematic study of sensitivity to SAE quality is beyond the present scope. Single layer and direction: we intervene at one mid-network layer along a single direction per target, following established practice; multi-layer, multi-direction, and composed-behavior steering, together with the allocation of a budget across them, define a natural extension of the budgeted-allocation view that we do not pursue here. Targets and language: we study three targets (toxicity, fear, and sadness) selected for their cleanly separable, classifier-scorable endpoints on English text, which establishes the effect under controlled conditions; behaviors without a reliable automatic scorer and non-English settings raise distinct calibration and evaluation considerations that we do not address. Automatic evaluation: behavioral efficacy and instruction following are assessed with automatic classifiers and two independent LLM judges under position randomization, and we report the resulting agreement; this protocol reduces but does not fully remove evaluator bias, and a human evaluation lies outside the current study. Scope of the analysis: our stability result characterizes the closed loop through a first-order linearization around the interior operating point, where the calibrated budget keeps the steering probability away from its bounds; it is intended to explain why accumulated feedback allocates a scarce budget effectively, and we do not claim worst-case guarantees over the full nonlinear decoding process. Broader impact: activation steering can suppress harmful behavior and is inherently dual-use; our contribution renders the intervention sparser and more controllable, and responsible deployment remains contingent on the chosen target direction.

Steering target: sadnessPrompt: “Nothing but net. He got a”
MethodppGenerated continuation
CFP (ours)0.5lot of help from his teammates, though. The 2018-19 season is over for the University at Albany men’s basketball team, and I’m not sure if it was more disappointing or just sad to see. . .
CLABS (ours)0.5lot of them. The 2018-19 season was the first time in my life that I didn’t watch every single game, and it’s not because there weren’t any good games to miss. . .
Full-Steer1.0lot of them. I was in the middle of writing this post when I saw that my friend. . . had written about how I am not good at being alone. It’s true.
No-Steer0lot of help from his teammates, though. The 2018-19 NBA season is officially underway and the Boston Celtics are off to an impressive start with their first win. . .
Promptingn/aperfect score on his math test, which was 100%. His parents were very proud of him.
SIA-ER0.5lot of them. I was in the middle of writing this post when I saw that my friend. . . posted about how much I suck at being a human on Facebook. . .
SIA-PE0.5lot of help from the refs, though. I’m not sure if I should be more upset that he was able to get away with it or that they didn’t even try to stop him. . .
SRE0.5lot of them. The 2018-19 season is over, and the New York Knicks are in last place again. . . this was an awful year for the team’s fans.
STS0.5lot of them. I was in the stands at the 2018 NCAA Men’s Basketball Tournament when I saw it happen: a player on my team who had been struggling all game. . .

Table 8. Sadness steering. CFP and CLABS methods successfully steer the response to reflect sadness, whereas Full-Steer abandons the topic for personal loneliness, and Prompting inverts the target into a cheerful output.

References

  1. [1]Anthropic. 2026. Claude Sonnet 4.6. https://www.anthropic.com/claude/sonnet. Large language model. Accessed 2026-07-22.
  2. [2]Åström, K. J.; and Wittenmark, B. 1997. Computer-Controlled Systems: Theory and Design. Prentice Hall, 3rd edition.
  3. [3]Bricken, T.; Templeton, A.; Batson, J.; Chen, B.; Jermyn, A.; et al. 2023. Towards Monosemanticity: Decomposing Language Models with Dictionary Learning. Transformer Circuits Thread.
  4. [4]Cunningham, H.; Ewart, A.; Riggs, L.; Huben, R.; and Sharkey, L. 2024. Sparse Autoencoders Find Highly Interpretable Features in Language Models. In ICLR.arxiv.org/abs/2309.08600
  5. [6]De Langis, K.; Koo, R.; and Kang, D. 2024. Dynamic Multi-Reward Weighting for Multi-Style Controllable Generation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing.
  6. [7]Demszky, D.; Movshovitz-Attias, D.; Ko, J.; Cowen, A.; Nemade, G.; and Ravi, S. 2020. GoEmotions: A Dataset of Fine-Grained Emotions. In ACL.arxiv.org/abs/2005.00547
  7. [8]Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Yang, A.; Fan, A.; et al. 2024. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783.arxiv.org/abs/2407.21783
  8. [9]Dubois, Y.; Li, X.; Taori, R.; Zhang, T.; et al. 2024. Alpaca-Farm: A Simulation Framework for Methods that Learn from Human Feedback. NeurIPS.
  9. [10]Ekman, P. 1992. An Argument for Basic Emotions. Cognition & Emotion, 6(3-4): 169–200.DOI
  10. [11]Eshraghi, N.; Gondara, L.; Huang, Y.; Suresh, S.; Teran, L.; Pradeep, J.; Xu, X.; and Chevalier, F. 2026. A Coin Flip Per Token: Bernoulli Sparse Steering of Large Language Models. arXiv preprint arXiv:2607.05615.arxiv.org/abs/2607.05615
  11. [12]Fayyaz, M.; Modarressi, A.; Deilamsalehy, H.; Dernoncourt, F.; Rossi, R.; Bui, T.; Schütze, H.; and Peng, N. 2026. Steering MoE LLMs via Expert (De)Activation. In ICLR.arxiv.org/abs/2509.09660
  12. [13]Feng, Z.; Li, T.; Zhu, Z.; Zhou, H.; Qian, J.; Zhang, L.; Chua, J. J. D.; Mak, L. O.; and Ng, G. W. 2026. Fine-Grained Activation Steering: Steering Less, Achieving More. In ICLR.arxiv.org/abs/2602.04428
  13. [14]Ferrando, A.; Suau, X.; González, J.; and Rodriguez, P. 2025. Dynamically Scaled Activation Steering. arXiv preprint arXiv:2512.03661.arxiv.org/abs/2512.03661
  14. [15]Fierro, C.; and Roger, F. 2026. Steering Language Models with Weight Arithmetic. In ICLR.arxiv.org/abs/2511.05408
  15. [16]Gan, J.; Cheng, Z.; Jiang, Z.; Wang, C.; Yin, Y.; Luo, X.; Fu, Y.; and Gu, Q. 2025. Steering When Necessary: Flexible Steering Large Language Models with Backtracking. arXiv preprint arXiv:2508.17621.arxiv.org/abs/2508.17621
  16. [17]Gehman, S.; Gururangan, S.; Sap, M.; Choi, Y.; and Smith, N. A. 2020. RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models. In Findings of EMNLP.
  17. [18]Goyal, A.; Rathi, V.; Yeh, W.; Wang, Y.; Chen, Y.; and Sundaram, H. 2025. Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.arxiv.org/abs/2505.14536
  18. [19]Hartmann, J.; Heitmann, M.; Siebert, C.; and Schamp, C. 2023. More than a Feeling: Accuracy and Application of Sentiment Analysis. International Journal of Research in Marketing, 40: 75–87.DOI
  19. [20]Hellerstein, J. L.; Diao, Y.; Parekh, S.; and Tilbury, D. M. 2004. Feedback Control of Computing Systems. Hoboken, NJ: John Wiley & Sons.DOI
  20. [22]Keskar, N. S.; McCann, B.; Varshney, L. R.; Xiong, C.; and Socher, R. 2019. CTRL: A Conditional Transformer Language Model for Controllable Generation. arXiv preprint arXiv:1909.05858.arxiv.org/abs/1909.05858
  21. [23]Labonne, M. 2024. Uncensor Any LLM with Abliteration. https://huggingface.co/blog/mlabonne/abliteration.
  22. [24]Lee, B. W.; Padhi, I.; Natesan Ramamurthy, K.; Mie hling, E.; Dognin, P.; Nagireddy, M.; and Durandhar, A. 2025. Programming refusal with conditional activation steering. In International conference on learning representations.arxiv.org/abs/2409.05907
  23. [27]Lieberum, T.; Rajamanoharan, S.; Conmy, A.; Smith, L.; Sonnerat, N.; Varma, V.; Kramar, J.; Dragan, A.; Shah, R.; and Nanda, N. 2024. Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP.
  24. [28]Liu, A.; Sap, M.; Lu, X.; Swayamdipta, S.; Bhagavatula, C.; Smith, N. A.; and Choi, Y. 2021. DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts. In ACL.arxiv.org/abs/2105.03023
  25. [29]Liu, S.; Ye, H.; Xing, L.; and Zou, J. 2024. In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering. In ICML.arxiv.org/abs/2311.06668
  26. [30]Marks, S.; and Tegmark, M. 2024. The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations for True/False Datasets. arXiv preprint arXiv:2310.06824.arxiv.org/abs/2310.06824
  27. [31]Nguyen, D. V.; Vu, H. M.; Pham, N. Y.; Zhang, L.; and Nguyen, T. M. 2026. Activation Steering with a Feedback Controller. In ICLR.arxiv.org/abs/2510.04309
  28. [32]OpenAI. 2025. GPT-5 mini. https://platform.openai.com/docs/models/gpt-5-mini. Accessed: 2026-07-21.
  29. [33]Rimsky, N.; Gabrieli, N.; Schulz, J.; Tong, M.; Hubinger, E.; and Turner, A. M. 2024. Steering Llama 2 via Contrastive Activation Addition. In ACL.arxiv.org/abs/2312.06681
  30. [34]Soo, S.; Guang, C.; Teng, W.; Balaganesh, C.; Guoxian, T.; and Ming, Y. 2025. Interpretable Steering of Large Language Models with Feature Guided Activation Additions. Workshop on Building Trust in ICLR.arxiv.org/abs/2501.09929
  31. [35]Stolfo, A.; Balachandran, V.; Yousefi, S.; Horvitz, E.; and Nushi, B. 2025. Improving Instruction-Following in Language Models through Activation Steering. In ICLR.
  32. [36]Subramani, N.; Suresh, N.; and Peters, M. E. 2022. Extracting Latent Steering Vectors from Pretrained Language Models. In Findings of ACL.
  33. [37]Team, G.; Riviere, M.; Pathak, S.; Sessa, P. G.; Hardin, C.; Bhupatiraju, S.; Hussenot, L.; Mesnard, T.; Shahriari, B.; Ramé, A.; et al. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118.
  34. [38]Templeton, A.; Conerly, T.; Marcus, J.; Batson, J.; Rosen, B.; et al. 2024. Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread.
  35. [39]Turner, A. M.; Thiergart, L.; Udell, D.; Leech, G.; Mini, U.; and MacDiarmid, M. 2023. Activation Addition: Steering Language Models Without Optimization. arXiv preprint arXiv:2308.10248.
  36. [41]Yang, W.; Qiu, X.; Yu, L.; Zhang, Y.; Yang, A.; Kokhlikyan, N.; Cancedda, N.; and Garcia-Olano, D. 2025. Hallucination Reduction with CASAL: Contrastive Activation Steering for Amortized Learning. In ICLR.arxiv.org/abs/2510.02324
  37. [42]Zou, A.; Phan, L.; Chen, S.; Campbell, J.; Guo, P.; Ren, R.; Pan, A.; Yin, X.; Mazeika, M.; et al. 2023. Representation Engineering: A Top-Down Approach to AI Transparency. arXiv preprint arXiv:2310.01405.arxiv.org/abs/2310.01405

Paper details

vlnc:W_rsfavsbn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents