Steering-informed Top-k in Vision Models

• Steering-informed Top-k is a hybrid explanation method combining causal steering interventions with top-k activation evidence from sparse autoencoder features in vision models.
• It mitigates the limitations of purely correlational or causal methods by grounding interventions in natural top-activating images and refining language model explanations.
• Empirical evaluations indicate that this hybrid approach consistently outperforms standard top-k techniques, especially for subtle and context-dependent visual features.

Searching arXiv for the cited papers to ground the article in current metadata and related work. {"query":"arXiv (Ferrando et al., 23 Mar 2026) LLMs Can Explain Visual Features via Steering", "max_results": 5} {"query":"arXiv (Chatzoudis et al., 2 Jun 2025) Visual Sparse Steering Improving Zero-shot Image Classification with Sparsity Guided Steering Vectors", "max_results": 5} {"query":"arXiv (Kariyappa et al., 2023) SHAP@k Efficient and Probably Approximately Correct Identification of Top-k Features", "max_results": 5} {"query":"arXiv (Xie et al., 2020) Differentiable Top-k Operator with Optimal Transport", "max_results": 5} Steering-informed Top-$k$ is a hybrid explanation procedure for sparse autoencoder (SAE) features in vision models that combines two distinct interpretability signals: top-$k$ activation evidence from natural images and steering-based causal interventions in the model’s internal representation. In the formulation introduced in "LLMs Can Explain Visual Features via Steering" (Ferrando et al., 23 Mar 2026), an SAE is trained on a vision encoder, individual SAE features are interpreted by injecting their decoder directions into the forward pass, and a LLM is prompted to explain what the model “sees” after the intervention. Steering-informed Top-$k$ then uses those steering outputs to ground and refine explanations derived from the dataset images where the same feature activates most strongly. The method is presented as a way to combine causal and correlational evidence, and as a route to state-of-the-art explanation quality without additional computational cost (Ferrando et al., 23 Mar 2026).

1. Definition and motivation

In this setting, a feature is a latent coordinate of an SAE trained on internal representations of a vision encoder. Steering-informed Top-$k$ is defined as a hybrid explanation method for such features. For a feature $i$, the procedure uses steering by injecting the $i$-th row of the SAE decoder into the model’s forward pass, and uses Top-$k$ by retrieving the dataset images or patches where the feature’s activation is highest. The steering step informs how the top-$k$ evidence is interpreted, while the top-$k$ images ground the steering-based description in natural data (Ferrando et al., 23 Mar 2026).

The motivation follows from complementary failure modes of the two component methods. Top-$k$ explanations are correlational: they summarize what highly activating examples have in common, but high activation alone does not guarantee that the resulting textual explanation is causally linked to the feature. This can be confounded by co-occurring visual patterns, can require scanning large datasets, and can be especially weak for subtle background or contextual features. Steering-based explanations are causal in the sense that they intervene directly on the internal representation, but the intervention can be off-manifold, particularly when applied from a blank or null image, and the resulting description can be influenced by language priors rather than solely by the feature’s effect (Ferrando et al., 23 Mar 2026).

The hybrid method is therefore motivated by two reciprocal goals: to add causal grounding to Top-$k$0 and to add data grounding to steering. The paper explicitly states that Steering-informed Top-$k$1 “reduces this issue by grounding the intervention in natural top-activating images” and “consistently improves over Top-k alone” (Ferrando et al., 23 Mar 2026). A plausible implication is that the method is best understood not as a replacement for example-based interpretability, but as a mechanism for reconciling causal intervention with natural-image evidence.

Component	Evidence source	Main weakness in isolation
Top-$k$2 explanations	Highest-activating dataset images or patches	Correlational and dataset-expensive
Steering-based explanations	Injected SAE feature direction plus language-model description	Off-manifold and potentially ungrounded
Steering-informed Top-$k$3	Steering output plus top-$k$4 natural examples	Combination rule not fully specified in the excerpt

2. SAE feature model and representational basis

The hybrid procedure assumes a vision encoder with an internal representation $k$5 and an SAE with encoder $k$6 and decoder $k$7. The representation is approximated as a linear combination of feature-specific decoder directions: $k$8 Using standard notation consistent with the provided reconstruction,

$k$9

The text further notes the standard SAE objective

$k$0

while also stating that the exact form of the loss, the value of $k$1, and the encoder architecture are not specified in the excerpt (Ferrando et al., 23 Mar 2026).

Within this representation, feature $k$2 corresponds to the decoder row $k$3 together with its activation $k$4. The decoder row defines a direction in representation space, and the activation specifies the coefficient with which that direction contributes to reconstruction. This matters for Steering-informed Top-$k$5 because both constituent explanation modes operate on the same SAE features. The feature extraction stage is common; only the interpretive strategy differs (Ferrando et al., 23 Mar 2026).

The steering intervention is described by adding the decoder row to a baseline representation. In the notation reconstructed from the excerpt,

$k$6

The provided text explicitly mentions “adding the $k$7-th row of $k$8” and notes that a scaling coefficient $k$9, if used, is not specified. The steered representation is then propagated through the downstream model so that the LLM can describe the induced visual content (Ferrando et al., 23 Mar 2026).

3. Hybrid explanation procedure

The overall pipeline begins with SAE feature extraction on a chosen layer of the vision encoder. For each image $k$0 in a dataset $k$1, the model computes a layer representation $k$2 and sparse activations $k$3. For feature $k$4, the standard Top-$k$5 baseline retrieves the images or patches where $k$6 is largest; this is the usual top-activation set for that feature (Ferrando et al., 23 Mar 2026).

The steering branch uses a neutral baseline representation, described in the excerpt as a “blank image” or “null input.” For a given feature $k$7, the method injects the corresponding decoder direction into this baseline representation and then prompts a LLM with a simple captioning-style instruction to explain what the model “sees.” The text notes that if the prompt is kept without steering, the model describes a blank image, which serves as a control. The exact prompt is said to appear in Appendix G, but is not reproduced in the excerpt (Ferrando et al., 23 Mar 2026).

Steering-informed Top-$k$8 then combines the steering-generated textual explanation with the natural top-$k$9 examples. The excerpt does not provide formal pseudocode or a scoring equation for the combination step, but it describes the conceptual role clearly: steering provides a causal, feature-centric description, while Top-$i$0 provides natural visual evidence. The hybrid uses the former to interpret the latter and the latter to validate or refine the former (Ferrando et al., 23 Mar 2026).

A high-level procedural reconstruction consistent with the supplied material is:

1. Train an SAE on a layer representation of a vision encoder.
2. For each feature $i$1, collect top-$i$2 images where $i$3 is highest.
3. Construct a steered representation by adding $i$4 to a blank-image representation.
4. Ask the LLM to describe what is in the resulting “image.”
5. Use the steering description together with the top-$i$5 images to form the final explanation.

The excerpt also states that the system uses a rank-based voting strategy across metrics to select the best explanation per feature, together with minimum quality thresholds, but does not specify the exact thresholds or aggregation rule (Ferrando et al., 23 Mar 2026). This suggests that Steering-informed Top-$i$6 is not merely a presentation format; it is also embedded in an evaluation-and-selection pipeline.

4. Empirical behavior and evaluation

The paper reports that steering offers a scalable alternative that complements traditional approaches based on input examples, and that explanation quality improves consistently with the scale of the LLM (Ferrando et al., 23 Mar 2026). Within that framing, Steering-informed Top-$i$7 is presented as the strongest configuration: it combines causal intervention with top-activation evidence and is described as yielding state-of-the-art explanation quality without additional computational cost (Ferrando et al., 23 Mar 2026).

The supplied material identifies three families of evaluation signals: simulation-based, CLIP-based alignment, and synthetic-image-based metrics. It also notes that these metrics are designed to penalize explanations that are not causally linked to the target feature. Steering alone is said to achieve non-trivial performance across multiple metrics, despite the off-manifold concern, while the hybrid method “consistently improves over Top-k alone” (Ferrando et al., 23 Mar 2026).

The excerpt does not reproduce concrete numerical tables, metric formulas, or hyperparameter values such as the top-$i$8 size, steering strength, or language-model decoding parameters. It does, however, record two substantive qualitative findings. First, top-$i$9-only methods remain a strong baseline but are incomplete. Second, steering is “particularly effective” for background features, suggesting that causal intervention can isolate contextual structure that is hard to recover from example inspection alone (Ferrando et al., 23 Mar 2026).

A plausible interpretation is that the hybrid is especially valuable when a feature’s natural exemplars are visually cluttered or when the relevant pattern is consistently present but not visually dominant. In such cases, steering-based text can supply a feature-centric hypothesis, and the top-$i$0 images can test whether that hypothesis is realized in actual data.

5. Conceptual interpretation and common misconceptions

The central conceptual distinction is between causal and correlational evidence. Top-$i$1 examples answer a question of the form: On which inputs does this feature appear most strongly? Steering answers a different question: What does this feature do when it is directly injected into the model? Steering-informed Top-$i$2 is an attempt to unify those two queries within one explanation workflow (Ferrando et al., 23 Mar 2026).

One common misconception is to treat top-$i$3 examples as if they were already causal evidence. The paper’s framing rejects that equivalence. Top-$i$4 reveals co-occurrence under the data distribution, but a pattern that repeatedly appears in high-activation images may still be incidental to the feature’s actual causal role. Another misconception is to treat steering outputs as intrinsically faithful because they arise from intervention. The paper explicitly notes that steering from a blank image can create unnatural activations, so causal status alone does not guarantee ecological validity (Ferrando et al., 23 Mar 2026).

The hybrid construction is designed precisely around that tension. The steering branch risks off-manifold artifacts; the top-$i$5 branch risks ambiguity and confounding. By grounding interventions in natural top-activating images, Steering-informed Top-$i$6 tries to retain the causal advantage of steering while reducing its dependence on artificial baselines. This suggests a broader interpretability principle: internal interventions become more credible when checked against naturally occurring evidence, and natural examples become more informative when paired with a feature-specific causal probe.

The paper also situates the method near “causal self explanation” work in LLMs, in the sense that the model’s own internal representations and controlled interventions are used to generate explanations for learned features (Ferrando et al., 23 Mar 2026). The excerpt does not elaborate this connection in detail, but the analogy clarifies why the method is distinctive: explanation is not derived solely from correlated examples or external labels, but from the model’s response to a targeted change in its internal state.

6. Related uses of “steering” and “top-$i$7” in adjacent literature

The phrase “steering-informed Top-$i$8” is most specifically associated with the hybrid explanation procedure of (Ferrando et al., 23 Mar 2026), but related work uses the same ingredients—sparse features, top-$i$9 selection, and steering—in different configurations. In "Visual Sparse Steering: Improving Zero-shot Image Classification with Sparsity Guided Steering Vectors" (Chatzoudis et al., 2 Jun 2025), a top-$k$0 SAE is used not to explain features but to determine which sparse concepts are active for a given image and therefore eligible to contribute to a steering vector. There, the top-$k$1 operator acts as a steering gate: only the active sparse units are amplified, decoded, and used to modify the CLIP embedding at test time (Chatzoudis et al., 2 Jun 2025).

That usage is methodologically adjacent but substantively distinct. In (Chatzoudis et al., 2 Jun 2025), steering changes downstream zero-shot classification behavior; in (Ferrando et al., 23 Mar 2026), steering elicits textual descriptions of learned visual features. The former is an intervention for performance modulation, the latter an intervention for interpretability. The shared dependence on sparse features and top-$k$2 selection suggests a family resemblance, but the objectives are different (Chatzoudis et al., 2 Jun 2025).

Other top-$k$3 literatures provide still different meanings. "SHAP@k: Efficient and Probably Approximately Correct (PAC) Identification of Top-k Features" (Kariyappa et al., 2023) treats top-$k$4 as a sample-efficient identification problem under uncertainty, using confidence intervals and bandit-style sampling to identify the $k$5 largest SHAP values. "Differentiable Top-k Operator with Optimal Transport" (Xie et al., 2020) treats top-$k$6 as a differentiable approximation problem, replacing the discrete selection with an entropic optimal transport relaxation. These works concern adaptive sampling and differentiable optimization, respectively, rather than SAE-feature explanation. Their relevance is therefore conceptual rather than definitional: they illustrate that “top-$k$7” can be coupled with steering either as uncertainty-aware allocation (Kariyappa et al., 2023) or as gradient-based control (Xie et al., 2020), even though neither paper introduces the hybrid explanation protocol of (Ferrando et al., 23 Mar 2026).

7. Limitations and unresolved questions

Several limitations are explicit in the supplied material. First, language-model hallucination remains a live concern: the LLM may produce plausible but inaccurate explanations, even when prompted from an intervention. The mitigation described in the excerpt is reliance on multiple metrics that penalize explanations not causally linked to the target feature, but this does not eliminate the problem (Ferrando et al., 23 Mar 2026).

Second, off-manifold intervention is intrinsic to the blank-image steering setup. Steering-informed Top-$k$8 reduces the issue by grounding the explanation in natural top-activating images, but the basic intervention is still performed relative to an artificial baseline. The excerpt does not claim that the off-manifold problem is solved, only that it is mitigated (Ferrando et al., 23 Mar 2026).

Third, the evaluation pipeline depends on other learned models, including CLIP-based alignment and synthetic-image-based metrics, and the text notes that SAM-based metrics were adopted from previous work while also acknowledging their limitations. This creates the possibility of evaluator bias, even when multiple complementary metrics are used (Ferrando et al., 23 Mar 2026).

Fourth, the method still depends on a dataset for SAE training. The paper clarifies that steering does not require a dataset for the explanation step once the SAE has been trained, but the dataset cost remains part of the overall interpretability workflow and is shared with other feature-explanation methods in the same setup (Ferrando et al., 23 Mar 2026).

Open questions follow directly from these limitations. The excerpt points toward better control of off-manifold steering, tighter integration between language-model explanations and natural-image evidence, and more robust evaluation protocols. These directions are not specified in detail in the supplied text, so any stronger forecast would be speculative. What is clear is that Steering-informed Top-$k$9 is presented as a hybrid interpretability procedure whose main contribution is methodological: it treats causal intervention and top-activation exemplars as complementary sources of evidence for explaining SAE features in vision models (Ferrando et al., 23 Mar 2026).