Robust Concept Activation Vectors

Updated 2 February 2026

RCAV is a framework for concept-based interpretability that enhances traditional CAV methods by mitigating noise, sampling variability, and misalignment.
It incorporates techniques such as SuperActivator tail-thresholding, pattern-based direction recovery, and adversarial sampling to improve detection F1 scores and stability.
RCAV provides practical guidance for threshold selection and regularized training, demonstrating significant gains in robustness and interpretability across modalities.

Robust Concept Activation Vectors (RCAV) formalize a suite of methodologies for concept-based interpretability in deep neural networks, overcoming the noise, unreliability, and misalignment associated with standard CAV approaches. RCAV encompasses a diverse range of technical solutions—tail-focused detection (SuperActivator mechanism), sampling-theoretic variance control, pattern-based direction recovery, adversarial sampling, spatial alignment, and regularization—each grounded in empirical and theoretical advances across recent literature (Goldberg et al., 4 Dec 2025, Wenkmann et al., 28 Sep 2025, Pahde et al., 2022, Lysnæs-Larsen et al., 6 Nov 2025, Soni et al., 2020, Corbetta et al., 19 Aug 2025, Pfau et al., 2021).

1. Sources of Non-Robustness in Classical CAVs

Standard Concept Activation Vectors are typically constructed by separating representations of samples that do and do not contain the concept, using linear classifiers at a chosen layer. Let $f$ be the model mapping input $x$ to representations $z$ at layer $\ell$ , and $v_c$ a concept vector fitted by logistic regression or SVM. Two critical failure modes undermine reliability:

Distributional Overlap: The in-concept activation distribution $D_c^\mathrm{in} = \{ s_c(z): z \text{ from tokens containing }c \}$ and out-of-concept $D_c^\mathrm{out}$ overlap heavily, with many true-concept tokens indistinguishable from out-of-concept background. No unique global threshold separates the distributions; overlap mass $\int \min(p_\mathrm{in}(s), p_\mathrm{out}(s)) \, ds$ remains large (Goldberg et al., 4 Dec 2025).
Sampling Variability: CAVs depend on randomly sampled reference sets, inducing a variance that scales as $O(1/N)$ , where $N$ is reference set size. Small $N$ results in unstable vectors; variance only vanishes for large sample sizes (Wenkmann et al., 28 Sep 2025).
Directional Misalignment: Separability-based CAVs optimize for classification, not for signal fidelity; distractors unrelated to the concept can dominate, rotating the probe away from the true concept axis (Pahde et al., 2022, Lysnæs-Larsen et al., 6 Nov 2025).

2. Formal Mechanisms for Robustness

RCAV denotes any enhancement that enforces stability, fidelity, or noise-resilience in concept vector construction or usage.

2.1. SuperActivator Mechanism

The SuperActivator mechanism exploits the observation that reliable concept signals concentrate exclusively in the extreme high tail of $D_c^\mathrm{in}$ , beyond the highest quantiles of $D_c^\mathrm{out}$ (Goldberg et al., 4 Dec 2025). For concept $c$ :

Compute the empirical quantile $Q_{1-\delta}(S_\mathrm{val}^+(c))$ for sparsity fraction $\delta$ .
Define threshold $\tau_{c,\delta}^\mathrm{super}$ and select tokens $T_{c,\delta}^\mathrm{super} = \{ z : s_c(z) \geq \tau_{c,\delta}^\mathrm{super} \}$ .
At test time, predict $c$ present iff $\max_i s_c(z_i) \geq \tau_{c,\delta}^\mathrm{super}$ .

This tail thresholding delivers absolute F1 improvements of up to +14% across modalities and datasets, with optimal $\delta$ typically in 2–10% for images, 10–40% for text.

2.2. Sampling-Theoretic RCAV Construction

RCAVs can be computed to achieve bounded variance by estimating the scale $a = \operatorname{tr}(\Sigma)$ via pilot runs and setting $N \geq a/\epsilon$ for target variance $\epsilon$ (Wenkmann et al., 28 Sep 2025). Regularization ( $\lambda$ in logistic or SVM losses) and averaging over several draws further reduce variance.

2.3. Pattern-based Direction Recovery

Rather than classification, pattern-based RCAVs solve a regression $p$ minimizing $\|A - t p^\top - b 1\|_2^2$ (where $A$ is activation matrix, $t$ concept label vector). Analytically, $h^\mathrm{pat} = (\operatorname{cov}[A, t])/\operatorname{var}(t) = \mathbb{E}_{t=+1} a - \mathbb{E}_{t=-1} a$ , yielding a direction invariant to distractor noise and feature scaling (Pahde et al., 2022). These vectors are highly aligned with the ground-truth concept axis in experiments, and improve sensitivity testing and shortcut suppression.

2.4. Adversarial and Orthonormal Sampling

Adversarial Concept Activation Vectors (A-CAVs) augment references by adversarially perturbing positives and negatives along output gradients, magnifying margin and separability (Soni et al., 2020). Gram-Schmidt orthogonalization projects negatives outside the concept subspace, followed by averaging across multiple draws to further stabilize the RCAV direction, reducing recall variance by 3–7×.

2.5. Spatial and Translation-Invariant Probes

Spatial alignment is enforced via pixelwise losses against concept masks, while translation-invariance is introduced by restricting probe weights to be constant across spatial locations—yielding RCAVs robust to spatial perturbations and background variation (Lysnæs-Larsen et al., 6 Nov 2025).

3. RCAV Algorithms and Pseudocode

Different classes of RCAVs involve distinct extraction and application protocols:

3.1. SuperActivator Thresholding

for concept c in concepts:
    S = [s_c(z_i) for z_i in in-concept tokens]
    for δ in grid(0.01, ..., 0.5):
        τ = quantile(S, 1-δ)
        F1 = evaluate_detection_F1(τ)
    choose δ* maximizing F1
    record τ*, layer ℓ*
z = get_layer_embeddings(x, ℓ*)
if max(s_c(z)) >= τ*: predict presence

(Goldberg et al., 4 Dec 2025)

3.2. Sampling-Theoretic RCAV

Pilot: fit small- $N$ CAVs, estimate $a$ in variance decay $a/N + b$ .
Set $N=a/\epsilon$ , fit final RCAV.

(Wenkmann et al., 28 Sep 2025)

3.3. Pattern-Based RCAV

Extract $h^\mathrm{pat} = \mathbb{E}_{t=+1} a - \mathbb{E}_{t=-1} a$ , normalize, use as the concept vector for sensitivity or attribution.

(Pahde et al., 2022)

3.4. Adversarial + GS RCAV

Perturb samples by $\varepsilon \operatorname{sign}(\nabla_x \mathrm{logit})$ .
GS orthonormalize concept positives, project negatives, sample and train SVMs, average vectors.

(Soni et al., 2020)

3.5. Spatial RCAVs

Train probe with spatial mask loss.
For translation-invariance, restrict $\mathbf{v}$ to channel weights.

(Lysnæs-Larsen et al., 6 Nov 2025)

4. Empirical Results and Benchmarking

The effectiveness and robustness of RCAVs are established across a broad range of architectures and modalities:

Dataset/Task	Metric	Baseline (Prompt/TCAV)	RCAV/SuperActivator	Absolute Gain
MS-COCO, Vision	F1 Detection	0.69	0.83	+0.14
OpenSurfaces	F1 Detection	0.49	0.56	+0.07
Text Sarcasm	F1 Detection	0.74 (CLS)	0.87	+0.13
GoEmotions	F1	0.37	0.46	+0.09

SuperActivator tail-thresholding yields up to +0.13 F1 in concept attribution alignment (measured with LIME, SHAP, Grad-CAM), and reliably captures >90% of true concept samples above out-of-concept quantiles (Goldberg et al., 4 Dec 2025). Pattern-based RCAV directions yield 3× higher cosine similarity to true concept axes and perfect TCAV sensitivity on restricted tasks (Pahde et al., 2022). Adversarial sampling and averaging increase recall by up to 60 pp and diminish cross-seed variance by 3–7× (Soni et al., 2020). Spatial/aligned RCAVs improve hard accuracy, segmentation scores, and augmentation-robustness by 5–10 pp and +0.05–0.1 over baseline probes (Lysnæs-Larsen et al., 6 Nov 2025).

5. Theoretical Foundations of Robustness

Tail-Only Robustness: Sparse upper-tail activations are minimally contaminated by noise; inclusion of tokens outside the tail dilutes signal and lowers F1. Optimal selection is dataset- and modality-specific (Goldberg et al., 4 Dec 2025).
Variance Control: The $O(1/N)$ variance law enables explicit choice of sample size for desired reliability, with regularization and averaging offering further reductions (Wenkmann et al., 28 Sep 2025).
Pattern Extraction: The pattern vector, computed as covariance between activations and concept label, is invariant to uncorrelated distractors and stable under rescaling (Pahde et al., 2022).
Adversarial Margin: Input-space adversarial perturbation stretches separation in representation space, increasing recall and stability (Soni et al., 2020).
Spatial/Mask Losses: Pixelwise alignment losses mitigate probe reliance on spurious features, actionable via segmentation masks and translation-invariant channel pooling (Lysnæs-Larsen et al., 6 Nov 2025).

6. RCAV in Regularisation and Training

RCAVs are not only used in posthoc interpretability but also as direct regularizers in the training objective. The LCRReg framework synthesizes disentangled concept exemplars, learns concept vectors (pattern/SVM/CAR), and injects layerwise alignment or decision-boundary penalties into the model’s loss (Corbetta et al., 19 Aug 2025). This yields substantial gains in robustness to spurious correlations (+10–15 pp balanced accuracy OOD), improved OOD generalization (e.g., +1.1 pp Diabetic Retinopathy), and superior resistance compared to multitask learning, linear probes, and posthoc residual fits. Regularization is most effective with strong weights, static scheduling, and a single upfront RCAV computation per concept.

7. Practical Guidance and Limitations

Calibration: RCAV thresholding is always per-concept, per-layer; optimum tail fraction or sparsity should be selected via validation F1 (Goldberg et al., 4 Dec 2025).
Computation: Pilot variance estimation, adversarial perturbation, GS orthogonalization, and pattern computation are all tractable for reference sets $N\leq 300$ and latent dimensions $d\leq 10^{3}$ (Wenkmann et al., 28 Sep 2025, Soni et al., 2020).
Sample Selection: For stable RCAVs, concept sets should be diverse and disentangled; avoid up-sampling a small number of images (Wenkmann et al., 28 Sep 2025, Corbetta et al., 19 Aug 2025).
Limitations: Performance saturates for very large models, rare concepts may require expanded annotation, frequent recomputation of concept vectors introduces optimization noise, and some advanced variants require segmentation masks or extra processing (Corbetta et al., 19 Aug 2025).

References

"SuperActivators: Only the Tail of the Distribution Contains Reliable Concept Signals" (Goldberg et al., 4 Dec 2025)
"On the Variability of Concept Activation Vectors" (Wenkmann et al., 28 Sep 2025)
"Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence" (Pahde et al., 2022)
"Probing the Probes: Methods and Metrics for Concept Alignment" (Lysnæs-Larsen et al., 6 Nov 2025)
"Adversarial TCAV -- Robust and Effective Interpretation of Intermediate Layers in Neural Networks" (Soni et al., 2020)
"In-hoc Concept Representations to Regularise Deep Learning in Medical Imaging" (Corbetta et al., 19 Aug 2025)
"Robust Semantic Interpretability: Revisiting Concept Activation Vectors" (Pfau et al., 2021)