Feature-guided Activation Addition (FGAA)

• FGAA is an activation steering method that uses SAE latent space to isolate contrastive features for guiding model behavior.
• It employs density filtering, BOS filtering, and top‑k selection to ensure sparse, interpretable, and effective activation adjustments.
• Empirical results on Gemma-2 models demonstrate that FGAA improves the trade-off between steering precision and overall model coherence.

Feature-guided Activation Addition (FGAA) is an activation steering method for LLMs that moves steering from raw residual-stream coordinates into the latent feature space of a Sparse Autoencoder (SAE). It combines the contrastive supervision of desired versus undesired examples with SAE-targeted steering machinery: a contrastive target is first formed in SAE latent space, filtered to retain meaningful features, then mapped back into a residual-stream steering vector and added during generation with scale $\alpha$. The method was introduced as a more precise, interpretable, and reliable alternative to raw Contrastive Activation Addition (CAA), simple SAE decoder steering, and SAE-targeted steering, and was evaluated on Gemma-2-2B and Gemma-2-9B across nine steering tasks (Soo et al., 17 Jan 2025).

1. Definition and methodological lineage

FGAA belongs to the family of activation steering methods, in which a model’s hidden state is modified during inference by adding a steering vector at a chosen layer. Its immediate background is CAA, which constructs a steering vector by averaging residual-stream activation differences between paired positive and negative examples and then adds that vector at all token positions after the prompt during generation (Panickssery et al., 2023). FGAA retains the same basic intervention form, but changes the space in which the target direction is defined.

The central claim of FGAA is that steering directions computed directly in the residual stream are often opaque and imprecise. The paper characterizes standard activation steering directions as potentially containing broad, noisy, or spurious components that change behavior in unintended ways and degrade coherence. FGAA addresses this by inheriting contrastive supervision from desired versus undesired examples, as in CAA, while inheriting from SAE-targeted steering the idea that steering should be specified in interpretable SAE feature space and then translated back into model activation space using a learned linear effect approximator (Soo et al., 17 Jan 2025).

This places FGAA between two established poles. Relative to CAA, it preserves the desired-versus-undesired contrastive setup but applies it in an interpretable feature basis rather than raw activation coordinates. Relative to single-feature SAE steering, it does not assume that one latent cleanly captures the target concept. The paper therefore presents FGAA as a hybrid with new feature-selection machinery rather than as a minor variant of earlier activation addition methods (Soo et al., 17 Jan 2025).

2. Formal construction in SAE latent space

FGAA begins by defining a steering task and constructing two small sets of text examples: desired examples $X^+$ that express the target behavior and undesired examples $X^-$ that do not. For a chosen layer $l$, hidden activations $h_l(x)$ are collected from the base model. These activations are not contrasted directly. Instead, they are passed through an SAE encoder $f$, using mean SAE activations across tokens, and the contrastive difference is computed in latent feature space as

$$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$

This vector is intended to represent, at feature level, what distinguishes desired from undesired behavior (Soo et al., 17 Jan 2025).

FGAA then applies three successive filtering operations. The first is density filtering,

$$\mathbf{v}_{\text{filtered}}(i)= \begin{cases} 0 & \text{if } \rho(i)>\theta\ \mathbf{v}_{\text{diff}}(i) & \text{otherwise} \end{cases}$$

where $\rho(i)$ is the activation density of feature $i$, and the implementation uses $X^+$0. The second removes BOS-heavy features,

$X^+$1

The third is top-$X^+$2 selection,

$X^+$3

In practice, the final experiments set $X^+$4, because negative features often hurt performance and made results unstable, and sweep $X^+$5 with task-specific selection (Soo et al., 17 Jan 2025).

To convert the sparse SAE-space target into a residual-stream steering vector, FGAA uses a pretrained linear effect approximator. The paper describes the approximator as

$X^+$6

where $X^+$7 is the $X^+$8-dimensional steering vector, $X^+$9 is a $X^-$0 matrix, and $X^-$1 is a $X^-$2-dimensional bias. The paper also notes inconsistent notation between $X^-$3 and $X^-$4; the implementation formula given is

$X^-$5

The paper does not provide a more formal constrained optimization derivation beyond this formula. It further states that $X^-$6 is L1-normalized before this calculation and reports empirically that L1 normalization works better than L2 normalization for this step (Soo et al., 17 Jan 2025).

At inference time, the intervention is

$X^-$7

with all steering vectors L2-normalized before application. The vector is added to the residual stream at every token position during generation (Soo et al., 17 Jan 2025).

3. Feature filtering, interpretability, and geometry

The interpretability claim of FGAA is inseparable from its feature-selection pipeline. Density filtering is motivated by the observation that high-density features fire often across many inputs and are usually not very task-specific. BOS filtering removes a family of beginning-of-sequence-dominated features observed in Gemma SAEs that often introduce generation artifacts without contributing useful task-specific control. Top-$X^-$8 selection then makes the intervention explicitly sparse over SAE latents, rather than leaving it as a dense and uninterpretable activation difference (Soo et al., 17 Jan 2025).

Because the target is a sparse vector over SAE latents, individual selected features can be inspected. The paper reports that, for the Love task on Gemma-2-2B, selected positive features include “instances and expressions of love,” “expressions of love and emotional connections,” “phrases related to physical intimacy,” and “references to romantic relationships.” For Conspiracy on Gemma-2-9B, selected features include “references to government, intelligence agencies, and organized crime,” “terms related to political correctness and liberal ideologies,” and “references to crime, corruption, and political manipulation” (Soo et al., 17 Jan 2025).

FGAA’s geometry relative to adjacent methods is also informative. The paper reports low cosine similarity between FGAA and CAA directions, mostly below $X^-$9, indicating that feature-space filtering and effect-approximator reconstruction substantially change the final direction rather than merely denoising it. By contrast, FGAA vectors are highly aligned with SAE-TS vectors, with cosine similarity around $l$0 on average for both Gemma models. The paper interprets this as evidence that FGAA and SAE-TS often converge on similar directions, while FGAA tends to do better because it constructs those directions from multiple automatically selected features rather than a single manually chosen latent (Soo et al., 17 Jan 2025).

Two misconceptions are explicitly addressed by the reported results. First, FGAA is not simply “CAA in another basis”: the low cosine similarity with CAA argues against that reading. Second, FGAA is not intrinsically multi-feature: the paper emphasizes that it can degenerate to a single-feature solution when that is optimal, and in Gemma-2-9B Anger with $l$1, FGAA and SAE-TS become effectively identical. The appendix further notes that negative features often degrade performance because “anti-concepts” are semantically broad and heterogeneous; in the Praise example, positive features form a coherent cluster around encouragement and admiration, whereas negative features span criticism, crime, sarcasm, and fatigue (Soo et al., 17 Jan 2025).

4. Experimental configuration and evaluation protocol

The reported experiments use Gemma-2-2B and Gemma-2-9B. The SAE is a pretrained Gemma Scope SAE with 16,384 features on the residual stream at layer 12, and steering is also applied at layer 12 of the residual stream. The paper states that this layer was chosen primarily because public effect approximators were available only there and because training approximators for each layer is expensive. The effect approximators themselves are taken from Chalnev et al., the SAE-TS paper (Soo et al., 17 Jan 2025).

During evaluation, each steering vector is injected at every token position while generating 100 completions of length 33 tokens from the open-ended prompt "<bos>I think". Hyperparameters $l$2 are chosen via sweep; based on appendix analysis, the final setting uses $l$3 and tunes only $l$4. The steering scale $l$5 is varied across experiments; the paper notes that, for some example rollouts, the best scales were around 80 for Love and 120 for Conspiracy, though optimal scale is task-dependent (Soo et al., 17 Jan 2025).

Steering quality is evaluated on nine tasks: Anger, Christian Evangelist, Conspiracy, French, London, Love, Praise, Want to Die, and Wedding. Outputs are scored by gpt-4o-mini on two rescaled criteria in $l$6: a behavioral score $l$7, measuring whether the generated text exhibits the target behavior, and a coherence score $l$8, measuring whether the text remains grammatically and semantically coherent. These are combined as

$l$9

The use of the Behavioral-Coherence Score is intended to penalize the failure mode in which steering raises apparent behavioral alignment by producing incoherent output saturated with target-related words (Soo et al., 17 Jan 2025).

General capability retention is evaluated separately. The paper measures perplexity on 100 OpenWebText samples and benchmark performance on MMLU and MMLU-Pro using Gemma-2-2B with steering vectors from representative tasks, while varying steering scale to track degradation in language modeling and downstream capability (Soo et al., 17 Jan 2025).

5. Empirical performance and scaling trade-offs

FGAA is reported as strongest on Gemma-2-2B. Averaged across the nine tasks, mean BCS is $h_l(x)$0 for FGAA, compared with $h_l(x)$1 for SAE-TS, $h_l(x)$2 for CAA, and $h_l(x)$3 for single-feature SAE steering. FGAA is best on 8 of 9 tasks for the 2B model. The paper highlights particularly large gains on Praise ($h_l(x)$4 versus $h_l(x)$5 SAE, $h_l(x)$6 SAE-TS, and $h_l(x)$7 CAA), Love ($h_l(x)$8 versus $h_l(x)$9 SAE-TS), Wedding ($f$0 versus $f$1 SAE-TS), and Christian ($f$2 versus $f$3 SAE-TS and $f$4 CAA) (Soo et al., 17 Jan 2025).

On Gemma-2-9B, the pattern is more mixed. FGAA still has the best average BCS, $f$5, compared with $f$6 for SAE-TS, $f$7 for CAA, and $f$8 for SAE steering, but task wins are shared more evenly. FGAA wins French, London, Love, and Praise; CAA wins Anger, Christian, and Conspiracy; SAE or SAE-TS win Want to Die and Wedding. The paper interprets this as evidence that FGAA’s gains may not scale monotonically with model size and that effectiveness is architecture-dependent (Soo et al., 17 Jan 2025).

Capability trade-offs are reported as consistent across all steering methods. At low steering scales, roughly below $f$9–$$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$0, perplexity and benchmark performance remain close to baseline. Around scale $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$1 there is a visible inflection point where steering starts to significantly impair general capability. Between scales $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$2 and $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$3, MMLU and MMLU-Pro decline sharply, and beyond about $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$4 generations become highly incoherent for all methods. The paper therefore treats stronger steering as reliably coupled to a capability cost regardless of method. It also notes a qualitative difference: CAA, SAE, and SAE-TS sometimes show a slight benchmark bump at very low scales, possibly a beneficial-noise effect, whereas FGAA shows more stable behavior without that bump (Soo et al., 17 Jan 2025).

These findings give FGAA a specific empirical profile. It improves the steering/coherence trade-off on a broad task set, especially for Gemma-2-2B, but it does not eliminate the general capability degradation associated with large activation additions. The method’s main reported benefit is therefore selectivity and interpretability at moderate steering scales rather than unrestricted behavioral control (Soo et al., 17 Jan 2025).

6. Relation to adjacent activation-intervention methods

FGAA is most directly connected to CAA. CAA constructs a steering vector by averaging residual-stream activation differences between behavior-positive and behavior-negative examples at a chosen layer and adds that vector at every generated token position after the prompt (Panickssery et al., 2023). FGAA retains the same additive intervention structure, but replaces raw residual-space differencing with contrastive differencing in SAE latent space, followed by feature filtering and effect-approximator reconstruction (Soo et al., 17 Jan 2025).

A distinct neighboring approach is task-driven layerwise additive activation intervention. That method also has the operational form of activation addition, but its intervention vector $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$5 is learned by minimizing task loss on a small supervised dataset, with sparse group lasso regularization over activation coordinates and attention heads. It therefore uses supervisory task guidance rather than feature guidance, does not identify semantically interpretable latent features, and applies the learned vector to the last-token activation at a single selected layer (Nguyen et al., 10 Feb 2025). Relative to FGAA, it is closer to a learned activation-space adapter than to an interpretable feature-guided steering method.

Outside inference-time steering, SAE-guided activation regularization for continual learning reports that task-relevant representations are linearly separable in the SAE feature basis on TRACE and MedCL, with pooled ROC-AUC values of $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$6 and $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$7, while neuron-basis directions are near chance at $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$8 and $$\mathbf{v}_{\text{diff}} = \frac{1}{|X^+|}\sum_{x\in X^+} f(h_l(x)) - \frac{1}{|X^-|}\sum_{x\in X^-} f(h_l(x)).$$9 (Ning et al., 25 Jun 2026). This suggests that SAE coordinates may provide a more selective intervention basis than raw neurons or weights, which is consistent with FGAA’s decision to define steering targets in SAE latent space rather than directly in residual coordinates.

More broadly, concept-level guidance has been argued to preserve representational diversity better than indiscriminate coordinate-level alignment. Concept Contrast replaces element-wise feature contrast with contrast over concept activation vectors built from clustered neurons and reports consistent gains in domain generalization, together with increased neuron coverage and lower hyperspherical potential energy (Liu et al., 2022). A plausible implication is that FGAA’s sparse feature selection may be more robust than dense coordinate-wise steering precisely because it specifies control in a semantically structured basis rather than in raw activation entries.

7. Limitations, misconceptions, and open questions

The paper identifies several practical limitations. FGAA depends on SAE quality, task-dependent hyperparameters such as $$\mathbf{v}_{\text{filtered}}(i)= \begin{cases} 0 & \text{if } \rho(i)>\theta\ \mathbf{v}_{\text{diff}}(i) & \text{otherwise} \end{cases}$$0, and the availability of effect approximators. It is implemented only where public approximators are available, which in the reported study restricts steering to layer 12. The SAE itself is sourced from pretrained Gemma Scope releases rather than trained as part of the method, so the paper does not provide a full SAE training recipe (Soo et al., 17 Jan 2025).

A second limitation is incomplete robustness on larger models. Although FGAA has the best average BCS on Gemma-2-9B, CAA still wins several tasks, and the paper explicitly states that effectiveness is architecture-dependent. FGAA should therefore not be interpreted as uniformly dominating raw contrastive steering across scales or model families (Soo et al., 17 Jan 2025).

A third limitation concerns the meaning of “feature-guided.” In FGAA, “feature-guided” does not mean that steering is defined by an a priori semantic ontology, nor that all selected features are guaranteed monosemantic. It means that the intervention is explicitly specified by a selected subset of SAE features, rather than by a dense, uninterpretable activation-difference vector. The paper’s own interpretability examples support this claim, but they do not establish a theorem of feature purity or disentanglement (Soo et al., 17 Jan 2025).

The remaining open questions are mostly structural. The current formulation leaves feature selection somewhat empirical rather than fully principled; negative features are largely discarded because they often hurt performance; and large steering scales still degrade perplexity, MMLU, and MMLU-Pro. The results therefore support a narrower conclusion than unrestricted controllability: FGAA offers a semantically inspectable and usually more precise activation-steering interface when SAE features and effect approximators are available, but it does not remove the fundamental trade-off between targeted behavior induction and preservation of general model capability (Soo et al., 17 Jan 2025).