Activation Additions for LLM Steering

• Activation additions are inference-time interventions that add a constant steering vector to hidden activations, biasing LLM outputs toward desired behaviors without altering weights.
• They can be implemented in raw residual space or via interpretable SAE features, offering diverse control mechanisms with varying transparency and precision.
• Empirical studies reveal a trade-off between increased steering strength and overall model coherence, underscoring challenges in balancing targeted behavior with language competence.

Activation additions are inference-time interventions in LLMs in which a steering vector is added to hidden activations so as to bias generation toward a target behavior without modifying model weights. In the constant additive form commonly used as a baseline in recent steering work, the intervention is written as $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}$$, where $\mathbf{z}_{attr}$ is a steering direction and $\alpha$ is a scalar steering coefficient. Recent work treats this mechanism as both a practical control primitive and a limitation: raw residual-stream additions are lightweight and flexible, but they may be opaque, entangle unrelated behaviors, and apply an overly uniform intervention across tokens (Soo et al., 17 Jan 2025, Heyman et al., 5 May 2026).

1. Position within activation steering

Activation additions belong to the broader family of activation steering or activation engineering methods, which intervene directly on activations at inference time rather than relying on prompting alone or on parameter updates. This makes them attractive when fine-tuning is expensive and prompt-based control is weak or easy to bypass. Their basic operational premise is simple: a hidden state is displaced by an additive vector chosen to correlate with a desired attribute or behavior (Soo et al., 17 Jan 2025).

In contemporary formulations, the central simplifying assumption is constancy. The same steering vector is added with a single global coefficient across generated tokens. In the notation used to critique this family, the standard form is

$$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$

This representation captures the essential mechanism of activation additions, but it also exposes why the method is often described as coarse. A fixed vector and a fixed scalar do not adapt to token position, local context, or layerwise heterogeneity, even though later analyses argue that successful prompt steering is strongly token-dependent (Heyman et al., 5 May 2026).

A common misconception is that activation additions are equivalent to prompt steering in a more direct or distilled form. The recent literature does not support that equivalence. Rather, it treats activation additions as a useful but restricted approximation to a richer intervention process that prompting induces internally.

2. Construction of steering directions

The simplest steering vectors are built in raw activation space. Contrastive Activation Addition (CAA), as summarized in later work, constructs a steering vector from the average residual-stream difference between desired and undesired examples. In that formulation, the method asks which activation direction separates positive and negative examples, and then reuses that direction as an additive intervention during generation (Soo et al., 17 Jan 2025).

Feature-space variants replace raw residual coordinates with a Sparse Autoencoder (SAE) latent basis. In "Interpretable Steering of LLMs with Feature Guided Activation Additions" (Soo et al., 17 Jan 2025), the first stage is a contrastive difference computed in SAE space: $$\mathbf{v}_\text{diff} = \frac{1}{|X^+|}\sum_{x \in X^+}f(h_l(x)) - \frac{1}{|X^-|}\sum_{x \in X^-}f(h_l(x)).$$ Here $h_l(x)$ is the hidden state at layer $l$, and $f(h_l(x))$ is the SAE latent activation vector averaged across tokens.

The paper then applies three explicit filters. First, high-density features are removed using a threshold $\theta = 0.01$: $$\mathbf{v}_\text{filtered}(i) = \begin{cases} 0 & \text{if } \rho(i) > \theta \ \mathbf{v}_\text{diff}(i) & \text{otherwise.} \end{cases}$$ Second, beginning-of-sequence-dominated features are removed. Third, the method keeps the strongest positive and negative features,

$\mathbf{z}_{attr}$0

although the main experimental configuration later sets $\mathbf{z}_{attr}$1 because negative features tended to degrade performance.

The filtered latent target is then mapped back to model space through a pretrained linear effect approximator, described as

$\mathbf{z}_{attr}$2

The optimized steering vector is reported as

$\mathbf{z}_{attr}$3

with the caveat that the paper notes imperfect typesetting while making clear that the target SAE feature vector is transformed through a learned linear effect model. An implementation detail emphasized by the authors is that $\mathbf{z}_{attr}$4 is $\mathbf{z}_{attr}$5-normalized before this optimization. The final intervention reverts to the activation-addition pattern: $\mathbf{z}_{attr}$6

3. Interpretable feature-space variants

The recent literature places several steering methods on a spectrum from opaque residual-space addition to more interpretable feature-space control. The core distinctions can be summarized compactly.

Method	Steering space	Distinctive property
CAA	Residual stream	Average difference between desired and undesired examples
SAE feature steering	SAE feature space	Uses the decoder vector of a single relevant SAE feature
SAE-TS	SAE feature space with effect approximation	Sets one SAE feature as the only active feature in the target vector
FGAA	SAE feature space with filtering and optimization	Uses contrastive SAE features, feature filtering, and a linear effect approximator
PSR	Activation space with token-specific scaling	Learns a coefficient function from activations themselves

The interpretability claim for SAE-based variants is not merely rhetorical. The argument is that SAE features can often be auto-labeled with human-readable descriptions, which allows practitioners to inspect the contents of a steering vector, remove irrelevant components, and analyze unintended side effects. By contrast, a raw CAA direction lives in activation space without a comparable latent decomposition (Soo et al., 17 Jan 2025).

A notable empirical observation is that FGAA and SAE-TS frequently discover very similar directions. Cosine similarity between FGAA and SAE-TS steering vectors is often greater than $\mathbf{z}_{attr}$7, whereas CAA and direct SAE vectors are often below $\mathbf{z}_{attr}$8. This suggests that multi-feature selection and single-feature targeting can converge toward related steering directions when the relevant latent structure is clean, even though the construction pipelines differ (Soo et al., 17 Jan 2025).

4. Empirical behavior on Gemma models

The main controlled evaluation of feature-guided activation additions uses Gemma-2-2B and Gemma-2-9B, together with a Gemma Scope SAE containing 16,384 SAE features, the residual stream at layer 12, and pretrained effect approximators for layer 12. The steering tasks are Anger, Christian Evangelist, Conspiracy, French, London, Love, Praise, Want to Die, and Wedding. For generation, the experiments sample 100 steered completions, each 33 tokens long, from prompts beginning with "<bos>I think". All steering vectors are $\mathbf{z}_{attr}$9-normalized before application for fair comparison (Soo et al., 17 Jan 2025).

Evaluation uses GPT-based scoring on a $\alpha$0–$\alpha$1 scale rescaled to $\alpha$2, producing a behavior score $\alpha$3 and a coherence score $\alpha$4, with

$\alpha$5

The Behavioral-Coherence Score is the main task metric. Broader capability preservation is assessed with perplexity on OpenWebText, MMLU, and MMLU-Pro.

On Gemma-2-2B, FGAA is reported as the best overall method. It wins on 8 of 9 tasks. The average BCS values are 0.2201 for CAA, 0.1404 for SAE, 0.3650 for SAE-TS, and 0.4702 for FGAA. On Gemma-2-9B, the pattern is more mixed, but FGAA remains competitive overall, with average BCS values of 0.2729 for CAA, 0.2267 for SAE, 0.3467 for SAE-TS, and 0.3979 for FGAA. CAA sometimes performs better on sentiment-heavy tasks in the larger model (Soo et al., 17 Jan 2025).

The coherence claim is central. The paper argues that raw CAA can be too opaque, direct SAE feature steering can be too narrow or aggressive, and FGAA better balances target strength with grammatical and semantic plausibility. This does not imply that coherence is guaranteed; rather, it indicates that more structured latent-feature selection can reduce some of the side effects associated with simpler activation additions.

5. Steering scale and capability degradation

A recurring empirical regularity is the trade-off between steering strength and general model competence. In the Gemma evaluations, stronger steering eventually hurts general capability regardless of method. At low scales, specifically below $\alpha$6, all methods preserve performance fairly well. Around scale $\alpha$7–$\alpha$8, there is an inflection point at which degradation becomes noticeable. Between roughly $\alpha$9 and $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$0, MMLU and MMLU-Pro performance falls sharply, and at very high scales outputs become incoherent for all methods (Soo et al., 17 Jan 2025).

The same experiments report that perplexity rises with scale while benchmark accuracy drops with scale. The resulting picture is not method-specific but structural: increasing $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$1 yields stronger behavioral control at the cost of broader language competence. This directly counters the common assumption that activation additions can be made arbitrarily strong if the steering direction is good enough.

The authors also note a more subtle phenomenon. Some baseline methods show a slight low-scale improvement on MMLU-Pro, possibly due to a mild noise-like regularization effect, whereas FGAA does not exhibit this bump. Their interpretation is that FGAA is more targeted and less noisy. This suggests that more precise steering may remove accidental regularization effects even as it improves task-specific control (Soo et al., 17 Jan 2025).

6. Faithfulness to prompting and token-specific generalizations

A major critique of activation additions is developed in "Steer Like the LLM: Activation Steering that Mimics Prompting" (Heyman et al., 5 May 2026). That work formalizes prompt steering itself as an internal activation intervention: $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$2 It distinguishes the accumulative effect $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$3 from the local effect $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$4. Against that backdrop, constant activation steering is criticized as unfaithful because it assumes both a single direction and equal magnitude across tokens.

The required assumptions are stated explicitly. First, prompt steering at each token would need to satisfy

$$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$5

Second, it would need equal magnitude across token positions,

$$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$6

The paper reports that prompt steering violates the second assumption badly: intervention magnitude varies substantially across token positions and layers. On that view, activation additions are not wrong so much as overly uniform.

Prompt Steering Replacement (PSR) retains the notion of a steering vector but makes the coefficient token-specific: $$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$7 In experiments, the coefficient function is implemented as a single-layer probe with ReLU,

$$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$8

This preserves an additive steering direction while abandoning uniformity across tokens. The paper defines S-PSR, which intervenes at one layer, and A-PSR, which intervenes at all layers. Training uses either an MSE objective or a log-likelihood objective, together with a regularizer

$$\mathbf{A}_{y_i \mid AS} = \mathbf{A}_{y_i} + \alpha \mathbf{z}_{attr}.$$9

Empirically, PSR outperforms standard activation steering on Persona Vectors, often matches or exceeds prompt steering at comparable coherence, and A-PSR$$\mathbf{v}_\text{diff} = \frac{1}{|X^+|}\sum_{x \in X^+}f(h_l(x)) - \frac{1}{|X^-|}\sum_{x \in X^-}f(h_l(x)).$$0 sets a new state of the art on both AxBench subsets. On IFEval, however, prompt steering remains stronger in several cases, and the paper states that rank-1 PSR is not expressive enough for all instruction types. The broader conclusion is therefore conditional rather than absolute: activation additions are a useful starting point, but more faithful steering may require token-specific, layer-aware modulation rather than a single constant additive vector (Heyman et al., 5 May 2026).