Feature Guided Activation Addition

• Feature Guided Activation Addition is a paradigm that uses externally or internally derived features to adjust neural activations for interpretability and task adaptation.
• It implements methods such as direct additive steering, feature-conditioned gating, and structural network interventions to achieve precise activation control.
• Empirical evaluations show that FGAA improves model performance and efficiency in architectures like LLMs, CNNs, and vision transformers by guiding activation modifications.

Feature Guided Activation Addition (FGAA) encompasses a set of neural network intervention and architectural techniques in which additive or multiplicative modifications to activations are computed or selected according to features—either externally provided or internally discovered within the model itself. These approaches enable precise, interpretable, and task-adaptive manipulation of high-dimensional hidden states in a variety of modern architectures, including LLMs, convolutional neural networks (CNNs), and vision transformers (ViTs). The conceptual core is to leverage features, either from input data or learned representations, as guidance for constructing activation modifications that are targeted, efficient, or enable network growth without output-altering retraining.

1. Conceptual Principles of Feature Guided Activation Addition

Feature Guided Activation Addition formalizes the process of modifying neural activations using explicitly feature-driven, rather than purely parametric, interventions. There are three principal incarnations:

• Direct activation steering: An explicit additive vector is constructed by contrasting activations for positive and negative examples corresponding to a feature of interest, then injected at a chosen model layer to steer behavior. When the contrast and selection are done in a feature-interpretable (e.g., SAE) space, this yields enhanced interpretability and targeting as in FGAA (Soo et al., 17 Jan 2025), or in the canonical "activation addition" paradigm (Turner et al., 2023).
• Feature-conditioned gating/activation: Multiplicative refinement (attention-like gating or polynomial modulation) of activations, where the gating signal is computed from features or local statistics of the input (e.g., ATAC units (Dai et al., 2020), adaptive activation networks (Jang et al., 2018)).
• Feature-preserving structural network interventions: Introduction of new neurons or layers, with the guarantee—via specially constructed refinable, sum-to-identity activations—that the model's input–output map is unchanged, thus enabling width/depth increases guided by subdivision-theoretic features (López-Ureña, 2024).

The unifying theme is that features, whether externally sourced (e.g., sparse autoencoder basis, spatial heatmaps, or class-discriminativity) or internally constructed, govern the structure, targeting, and effect of the activation modification.

2. Mathematical Formulations and Variants

The FGAA paradigm admits diverse mathematical realizations, each tailored to specific architectures and intervention goals.

Direct Additive Steering in LLMs

• Contrastive Activation Addition (CAA): Given two prompt sets $X^+$, $X^-$ exemplifying a desired/non-desired feature, the steering vector at layer $l$ is

$$v_{\rm steer} = \alpha \Bigg( \frac{1}{|X^+|}\sum_{x\in X^+} h_l(x) - \frac{1}{|X^-|}\sum_{x\in X^-} h_l(x) \Bigg)$$

where $h_l(x)$ is the activation, and $\alpha$ scales the intervention (Turner et al., 2023).

• Sparse Autoencoder Feature Steering (SAE-TS, FGAA): In the latent SAE space, construct a sparsified vector $v_{\rm target}$ selecting top- $k$ features, then map back by a learned linear "effect approximator" $W$:

$v_{\rm opt} = W\,v_{\rm target}$

yielding a model-space steering vector of controlled semantic scope (Soo et al., 17 Jan 2025).

Multiplicative Feature-Guided Activation

• ATAC Unit:

$X^-$0

where $X^-$1 is a pointwise, channel-aware gating map computed from local context (Dai et al., 2020).

• Feature-guided polynomial activation:

For neuron $X^-$2 in layer $X^-$3:

$X^-$4

where $X^-$5 is a feature set (channel or spatial neighborhood), and $X^-$6 are outputs of a secondary network conditioned on $X^-$7 (Jang et al., 2018).

Refinable Sum-to-Identity Activations for Architectural Interventions

• A univariate function $X^-$8 satisfies:

$X^-$9

This allows neuron or layer addition with zero output change, with the new neurons/layers parameterized via shifted/scaled copies of $l$0 (López-Ureña, 2024).

3. Practical Realizations Across Architectures

Feature Guided Activation Addition methods have been instantiated in diverse neural architectures:

Architecture	FGAA Instantiation	Guidance Feature Type
LLMs (Transformers)	CAA, SAE-TS, FGAA	SAE features, prompt class
CNNs	ATAC, Adaptive Activation Networks	Local channel stats, neighborhoods
Vision Transformers	Convolutional activation guidance	wCAM (class activation maps)

• In LLMs, FGAA enables behavioral alignment (e.g., toxicity reduction, sentiment shift) by adding precision-crafted vectors during inference, outperforming existing contrastive methods in steering accuracy and coherence (Soo et al., 17 Jan 2025).
• In CNNs, spatial and channel attention gates computed from local features refine activations, improving accuracy and sample efficiency in convolutional backbones (Dai et al., 2020, Jang et al., 2018).
• In ViTs, tokens are split into "major" and "minor" based on class-activation maps; the latter are fused and selectively modulated using global information, reducing computational cost while preserving information flow (Zheng et al., 2023).

4. Empirical Evaluation and Performance Implications

FGAA methods consistently outperform non-feature-aware add/gate approaches on targeted tasks.

• Steering in LLMs: FGAA achieves mean Behavioral-Coherence Score (BCS) increases up to 0.4702 on Gemma-2-2B, compared with 0.2201 for direct CAA and 0.3650 for SAE-TS. Effects are robust across tasks and models, with interpretable feature selection mitigating coherence degradation at high intervention scales (Soo et al., 17 Jan 2025).
• CNN Activation Modulation: Replacing ReLUs with ATAC units in ResNet architectures yields consistent top-1 accuracy gains (e.g., ImageNet ResNet-50: 21.41% top-1 error for ATAC vs 23.30% for ReLU). Spatially-varying, feature-guided gating yields better performance than global gating (Dai et al., 2020).
• Parameter–Efficiency Tradeoffs: Adaptive activation networks provide superior accuracy improvements compared to equivalent-sized deep/wide CNN baselines, demonstrating that feature-guided activation addition can substitute for brute-force model scaling (Jang et al., 2018).
• Architectural Growth: Refinable, sum-to-identity activations allow seamless architecture expansion (wider or deeper) without changing the network's function, enabling efficient architectural search or model adaptation with zero retraining (López-Ureña, 2024).

5. Feature Guidance Mechanisms: Design and Analysis

The efficacy of FGAA depends critically on principled feature selection and guidance mechanisms:

• SAE/Latent-space feature filtering: Density-based and BOS-feature exclusion, with top-k selection, isolates features that are semantically aligned with steering goals and avoids general or spurious activations, improving both interpretability and robustness (Soo et al., 17 Jan 2025).
• Contrast Prompt Design in CAA/ActAdd: Careful selection of minimally contrasting prompt pairs ensures that the steering vector isolates the desired semantic direction and minimizes off-target behavior, a principle validated by qualitative and quantitative analyses in LLM intervention (Turner et al., 2023).
• Convolutional Class Activation for Token Selection: In vision tasks, pixelwise class activation maps derived from lightweight CNNs (e.g., weighted Grad-CAM++) robustly distinguish semantically salient from background tokens, enabling efficient token fusion and gating in ViT variants (Zheng et al., 2023).
• Activation Network Topology: Polynomial order in activation networks (K) and bottleneck reduction ratio (r) in ATAC units are critical architectural hyperparameters, with ablation studies demonstrating that moderate values (e.g., K=5, r=2) maximize accuracy gains while retaining efficiency (Dai et al., 2020, Jang et al., 2018).

6. Limitations, Trade-offs, and Universality

FGAA methods present several important trade-offs and open directions:

• Steering Magnitude vs. Model Coherence: Increasing the intervention scale improves task alignment but ultimately degrades language modeling (perplexity) or overall accuracy. FGAA shows more favorable scaling regimes than prior methods, but does not eliminate the inherent trade-off (Soo et al., 17 Jan 2025).
• Sparsity and Head-level Selection: In models with rich internal structure (e.g., transformer heads), head/group sparsity regularization is critical for targeting task-relevant components and achieving sample-efficient adaptation (Nguyen et al., 10 Feb 2025).
• Architectural Expansion Scope: Identity-summing activations enable expansion only within explicit interval constraints (input scaling), requiring attention to signal range for correctness (López-Ureña, 2024).
• Universality of Feature-Guided Circuits: In cognitive and in-context learning settings, feature-guided addition at intervention heads (or linear subspaces) encodes not only direct response but also error-correcting and compositional computation, as in few-shot addition circuits in Llama-3-8B (Hu et al., 8 May 2025).

7. Synthesis and Outlook

Feature Guided Activation Addition unifies a spectrum of modern neural network interventions, architectural search, and interpretability techniques under the principle of feature-aware, semantically-targeted activation modification. Its efficacy in both pre-trained model steering (LLMs, image models) and network architecture adaptation (CNNs, generic NNs) is empirically validated across multiple large-scale benchmarks and architectures. Key technical ingredients—contrast-based feature discovery, sparse autoencoder factorizations, local attention-gating, and sum-to-identity activation design—provide a mathematically grounded framework for precise, interpretable, and efficient model intervention. Open directions include deeper formalization of feature guidance selection algorithms, principled trade-off analyses for steering magnitude, and generalization of refinable activations for non-Euclidean or sequential model families (Soo et al., 17 Jan 2025, Turner et al., 2023, Dai et al., 2020, López-Ureña, 2024, Hu et al., 8 May 2025, Jang et al., 2018, Zheng et al., 2023, Nguyen et al., 10 Feb 2025).