Zero-Shot Faithful Textual Explanations via Directional-Derivative Influence on Predictions

Abstract: Zero-shot textual explanations aim to make image classifiers more transparent by probing their internal representations, without relying on task-specific supervision or LVLMs. However, existing methods often miss the features that truly drive the prediction, resulting in limited \textit{faithfulness} to the evidence underlying the model's decision. To address this, we propose FaithTrace. Motivated by the idea that faithful explanations should describe concepts that strongly influence the prediction, FaithTrace directly measures how much the representation induced by the explanation changes the class logit. We introduce an influence score, computed as the directional derivative of the class logit along the text-induced direction in the classifier's feature space, and use it as a proxy for faithfulness. Moreover, we extend this influence score into quantitative evaluation metrics, helping fill the gap in faithfulness evaluation for textual explanations. Experiments show that FaithTrace yields more faithful explanations than baselines, facilitating a more accurate understanding of the model. The code will be publicly released.

• The paper presents FaithTrace, a framework that produces faithful textual explanations by quantifying text-induced directional-derivative influence on classifier predictions.
• The method aligns the classifier's feature space with CLIP embeddings using a learned affine aligner, enabling objective evaluation with directional scores and influence curves.
• Experimental results across various models show significant improvements over baselines, enhancing interpretability and aiding in bias diagnosis in image classifiers.

Zero-Shot Faithful Textual Explanations via Directional-Derivative Influence on Predictions

Overview

The paper "Zero-Shot Faithful Textual Explanations via Directional-Derivative Influence on Predictions" (2605.16877) addresses the challenge of producing faithful, zero-shot textual explanations for image classifiers. Faithful explanations are critical for interpretability, as they should reflect the true decision-driving cues used by the classifier, rather than generic or superficial semantic matches. The principal innovation introduced is FaithTrace, a framework that leverages the directional derivative of the classifier's class logit in its internal feature space, induced by text-derived directions, as a quantitative proxy for explanation faithfulness.

Motivation and Limitations of Prior Work

Existing zero-shot explanation methods, such as CLIP-based similarity or concept image visualizations, predominantly rely on semantic or correlational similarity between text and image embeddings. While these methods can produce plausible descriptions, they often fall short in faithfulness—namely, their explanations may not correspond to features that actually influence the classifier's prediction. They fail to provide rigorous quantitative evaluation criteria for faithfulness and do not directly probe the classifier's internal representations, leading to explanations that are generic or that overlook subtle, decision-critical evidence.

FaithTrace Framework

FaithTrace operates on arbitrary pretrained image classifiers, without requiring retraining or additional supervision. The approach proceeds as follows:

• Feature Space Alignment: The classifier's feature space is mapped to the CLIP joint vision-language embedding space via a learned affine aligner. This enables comparison between the classifier's latent representation and text embeddings.
• Text-Induced Direction Computation: For a given textual candidate $t$, the direction vector $\hat{v}_t(x)$ in the classifier's feature space is determined as the gradient that increases CLIP similarity between the mapped visual feature and text embedding.
• Directional Derivative Influence Score: The influence score is computed as the directional derivative $\nabla_{z_f} g_c(z_f)^{\top} \hat{v}_t(x)$ of the class logit with respect to the text-induced direction. This quantifies how much the activation of the classifier shifts toward or away from a class when perturbed along the text direction.
• Explanation Selection: Candidate explanations (from a concept bank generated via LLM and VLM prompts) are ranked according to influence scores, and those with highest positive influence are selected.
• Quantitative Evaluation Metrics: Directional score (local influence) and influence curves (insertion/deletion impacts along the direction) are used as rigorous metrics for assessing faithfulness.

Influence-Based Faithfulness Metrics

The paper proposes two quantitative metrics:

• Directional Score: Measures the infinitesimal change in class logit along $\hat{v}_t(x)$, serving as a localized faithfulness proxy.
• Influence Curves: Aggregate logit changes across finite steps in both insertion (toward the explanation) and deletion (away). These curves are computed using margin-based confidence scores for improved comparability.

These metrics fill a gap in evaluation protocols by moving beyond correlational or human-masked region approaches and directly address whether the explanation direction is causally relevant to the classifier's prediction.

Experimental Results

Rigorous evaluation is conducted across multiple architectures (ResNet-18/50, DINO ResNet-50, ViT, DINO ViT-S/8) using ImageNet-1K. FaithTrace demonstrates consistent superiority over baselines (Random, Text-To-Concept, TEXTER) on both quantitative metrics and qualitative assessments.

• Directional Score: FaithTrace achieves significantly higher scores and lower negative rates across all models and top-$k$ retrieval scenarios.
• Influence Curves: Influence curve sums for insertion and deletion are markedly larger for FaithTrace, indicating stronger alignment between explanation and prediction-driving features.
• Qualitative Analysis: FaithTrace explanations are more class-specific and reflect visually-grounded attributes (e.g., plumage color, context cues), while baselines frequently yield generic or irrelevant descriptions.
• Feature Visualization: Perturbation along text-induced directions enhances corresponding visual evidence in feature visualizations, confirming the semantic validity of the direction vectors.
• Misclassification Diagnostics: Explanations produced by FaithTrace for misclassified examples reveal spurious correlations (e.g., background cues) and facilitate interpretability regarding dataset biases.

Implications and Future Directions

The FaithTrace methodology has several practical and theoretical implications:

• Faithful Interpretability: This approach enables post-hoc interpretability for any off-the-shelf image classifier, advancing trust and transparency in vision systems.
• Bias Diagnosis: Quantitative faithfulness metrics and targeted explanations allow for systematic identification of spurious dataset correlations and failure modes.
• Generalization: As the alignment mechanism is plug-and-play, it is applicable to both CNNs and Transformers, and across diverse classifier settings.
• Automated Evaluation: Influence-based metrics provide an objective foundation for evaluating explanation fidelity, crucial for deploying models in sensitive domains.
• Concept Bank Coverage: The finite nature of concept banks is a limiting factor. Future work may leverage richer, more adaptive concept generation, possibly via retrieval augmentation or cross-modal synthesis.
• Broader AI Applications: The principle of using directional derivatives for faithfulness assessment can inform textual explanations in other modalities (e.g., NLP tasks, multimodal reasoning), and extend to active feature probing for robust model debugging.

Conclusion

This paper presents FaithTrace, a zero-shot framework for generating and evaluating faithful textual explanations for image classifiers by measuring the directional-derivative influence of text directions on class logits. Empirical results highlight its superiority over existing methods in both quantitative and qualitative faithfulness. The influence-based paradigm sets a new standard for the interpretability of vision models, with implications for diagnostic transparency and robust deployment.