FaithTrace: Zero-Shot Explanation for Image Classifiers

• FaithTrace is a zero-shot explanation framework that maps classifier features into CLIP space to derive textual explanations via influence-based ranking.
• It operationalizes concept guidance by evaluating directional derivatives, ensuring that text cues causally support the image classifier's predictions.
• Evaluated on ImageNet with various architectures, FaithTrace outperforms baselines by reducing negative rates and effectively pinpointing key visual cues.

FaithTrace is a zero-shot method for generating faithful textual explanations of image classifier predictions. Its defining claim is that an explanation should not merely describe something visible in the image; it should describe a concept whose presence, in the classifier’s own internal representation, would actually push the model toward the predicted class [2605.16877]. In the setting of a pretrained image classifier (g \circ f : \mathcal{X} \to \mathbb{R}^C), FaithTrace operationalizes this idea by mapping classifier features into CLIP space, turning candidate texts into directions in the classifier feature space, and ranking those directions by the directional derivative of the target logit [2605.16877]. In adjacent literature, the name also functions as a broader research framing for faithfulness tracing: evaluating whether explanations, chains of thought, internal concept trajectories, or executable traces are genuinely coupled to the computation that yields an output rather than serving as plausible but potentially decorative rationalizations [2510.22362] [2605.24286] [2603.05167].

1. Definition and conceptual scope

FaithTrace, in its most specific sense, is a classifier-plug-in explanation framework for image classification. The classifier is written as
[
g \circ f : \mathcal{X} \to \mathbb{R}^C,
]
with (f) a vision encoder producing a feature vector
[
z_f = f(x)\in Z_f,
]
and (g) a classification head producing logits
[
y = g(z_f), \qquad y_c = g_c(z_f)
]
for class (c) [2605.16877]. The classifier has been trained only on images and labels; it has no language module and is not retrained for explanation [2605.16877].

The central distinction drawn by FaithTrace is between semantic relatedness and prediction support. Prior zero-shot textual explanation methods often rely on external vision-language similarity, typically CLIP-style cosine similarity, to retrieve descriptions that are semantically related to the image. FaithTrace argues that such descriptions may be plausible or generic while remaining irrelevant—or even counterproductive—to the classifier’s predicted logit [2605.16877]. Its alternative is influence-based validation: if the internal representation is nudged toward the concept encoded by a text explanation, the predicted-class logit should increase [2605.16877].

A plausible implication is that FaithTrace shifts explanation from correlational matching to local decision sensitivity. That broader shift aligns with a wider faithfulness literature in which explanation quality is not identified with plausibility alone. For chain-of-thought reasoning, related work defines faithful traces as those functionally integrated into the model’s computation and meaningfully covarying with the final answer, rather than decorative or post hoc verbalizations [2510.22362]. Another line formulates faithful CoT as mediated information flow through (P \rightarrow C \rightarrow A) rather than a direct (P \rightarrow A) shortcut [2605.24286]. A benchmark perspective further decomposes process faithfulness into causality and coverage, emphasizing that a trace can sound plausible while failing to support the answer structurally [2603.05167].

2. Mathematical formulation

FaithTrace first connects classifier feature space to language using a vision-language model, specifically CLIP. Following Text-To-Concept, it learns an affine aligner
[
h(z_f) = W z_f + b
]
that maps classifier features into CLIP’s image-embedding space [2605.16877]. The aligner is trained by minimizing squared error against CLIP image embeddings:
[
\min_{W,b} \frac{1}{|\mathcal{D}{\text{train}}|} \sum{x \in \mathcal{D}{\text{train}}} \left| W z_f+b-E{\text{img}}(x) \right|2^2,
]
where (E{\text{img}}) is the CLIP image encoder and (z_f=f(x)) [2605.16877].

A text explanation is represented by its normalized CLIP text embedding
[
\hat{z}{\text{text}} = \frac{E{\text{text}}(t)}{|E_{\text{text}}(t)|2},
]
and the aligned image representation is normalized as
[
\hat{z}_h = \frac{h(z_f)}{|h(z_f)|_2}.
]
Their similarity is
[
s(h(z_f), t) = \hat{z}_h^\top \hat{z}{\text{text}}
]
[2605.16877]. Unlike retrieval methods that directly rank text by this cosine similarity, FaithTrace uses the similarity only as a bridge to derive a text-induced direction in the classifier feature space [2605.16877].

For each candidate text (t), the method defines
[
v_t(x) = \frac{\partial s(h(z_f),t)}{\partial z_f}.
]
For an affine aligner (h(z_f)=Wz_f+b), the closed form is
[
v_t(x) = W^{\top}\frac{1}{|W z_f + b|2} \left( \hat z{\text{text}} - s(h(z_f),t)\,\frac{W z_f + b}{|W z_f + b|_2} \right).
]
In practice, FaithTrace uses the normalized direction
[
\hat{v}_t(x) = \frac{v_t(x)}{|v_t(x)|_2}
]
[2605.16877].

Its formal definition of faithfulness is based on the effect of a small perturbation of the classifier feature along the text-induced direction. For class (c), the influence score is
[
I_c(x; \hat{v}t(x)) = g_c\bigl(z_f+ \epsilon \hat{v}_t(x)\bigr)-g_c\bigl(z_f\bigr),
]
where (\epsilon>0) is a small step size [2605.16877]. Using first-order Taylor expansion,
[
g_c(z_f+\epsilon \hat{v}_t(x)) \approx g_c(z_f) + \epsilon \,\nabla{z_f} g_c(z_f)^{\top} \hat{v}t(x),
]
so
[
I_c(x; \hat{v}_t(x)) \approx \epsilon \,\nabla{z_f} g_c(z_f)^{\top} \hat{v}t(x).
]
The quantity
[
\nabla{z_f} g_c(z_f)^{\top} \hat{v}t(x)
]
is the directional derivative of the class logit at (z_f) along the text-induced direction:
[
\nabla{z_f} g_c(z_f)^{\top} \hat{v}t(x) = \lim{\delta\to 0}\frac{g_c(z_f+\delta \hat{v}_t(x)) - g_c(z_f)}{\delta}
]
[2605.16877]. This directional derivative is the mathematical core of FaithTrace.

The paper turns the same idea into evaluation metrics. The directional score is
[
\text{Directional Score} = \nabla_{z_f} g_c(z_f)^{\top} \hat v_{t'}(x),
]
and the influence curve defines insertion and deletion changes for a sequence of step sizes ({\alpha_k}{k=1}^K):
[
\Delta_c^{\mathrm{ins}}(\alpha_k) = g_c\bigl(z_f + \alpha_k \hat v{t'}(x)\bigr) - g_c(z_f),
]
[
\Delta_c^{\mathrm{del}}(\alpha_k) = g_c(z_f) - g_c\bigl(z_f - \alpha_k \hat v_{t'}(x)\bigr).
]
The step sizes are set relative to the feature norm,
[
\alpha_k = \rho_k \,|z_f|2,
]
and in practice the paper uses the margin
[
m_c(z_f) = g_c(z_f) - \max{j \neq c} g_j(z_f)
]
and the sigmoid-normalized confidence
[
q_c(z_f) = \sigma\bigl(m_c(z_f)\bigr)\in(0,1)
]
for the finite-step curves [2605.16877].

3. Pipeline and implementation

FaithTrace does not generate explanations freely; it selects them from a concept bank (\mathcal{B}(x,c)) of short candidate descriptions [2605.16877]. The bank combines class-level and image-specific concepts. Following TEXTER, the implementation uses an LLM prompted with the class label (c) to generate generic class-relevant short phrases, and a VLM prompted with ((x,c)) to generate visually grounded descriptions specific to the image [2605.16877]. The reported implementation uses GPT-3.5-turbo for the LLM and Qwen2.5-VL-7B-Instruct for the VLM. The bank contains 130 descriptions per prediction: 100 from the LLM and 30 from the VLM [2605.16877].

The full pipeline is explicit. Given image (x) and class (c), FaithTrace computes classifier feature (z_f=f(x)); aligns it to CLIP image space with (h(z_f)); generates the candidate bank (\mathcal{B}(x,c)); for each candidate (t\in\mathcal{B}(x,c)), computes the text-induced direction (v_t(x)=\partial s(h(z_f),t)/\partial z_f), normalizes it to (\hat v_t(x)), evaluates the influence via the directional derivative (\nabla_{z_f} g_c(z_f)^\top \hat v_t(x)), and ranks candidates by this influence. The top-(k) texts with the largest positive scores are returned as explanations [2605.16877].

The method is largely training-free with respect to the target classifier, but not entirely parameter-free overall. The classifier itself is fixed; no retraining or fine-tuning of the classifier is needed. The only learned component is the affine aligner (h), trained separately using image-only data and CLIP image embeddings. Required components are the target classifier (g\circ f), CLIP image and text encoders, the learned aligner (h), and external LLM/VLMs for candidate concept-bank construction [2605.16877].

The central design decision is where the ranking signal comes from. Text-To-Concept and related alignment methods rank by semantic similarity in a VLM space; TEXTER generates concept images first but still ranks by CLIP-space similarity to text. FaithTrace instead uses VLM alignment only to construct a text-induced direction, then asks whether that direction causally supports the classifier’s own prediction in feature space, as approximated by the directional derivative of the class logit [2605.16877].

4. Empirical performance and diagnostics

FaithTrace is evaluated on ImageNet-1K. The aligner is trained on 20% of the ImageNet training split, following Text-To-Concept’s setup, and explanation evaluation is conducted on the ImageNet-1K test set [2605.16877]. The paper tests across both CNNs and Transformers: ResNet-18, ResNet-50, DINO-pretrained ResNet-50, ViT, and DINO ViT-S/8. The CLIP space is defined by the ViT-B/16 CLIP image encoder [2605.16877].

The baselines are Random, Text-To-Concept, and TEXTER [2605.16877]. The main reported result is that FaithTrace substantially outperforms all baselines on all tested architectures and retrieval settings [2605.16877]. On the directional score metric in the top-1 setting, the paper reports the following mean directional score / negative rate pairs:

Model	FaithTrace mean directional score	Negative rate
ResNet-18	0.161	0.037
ResNet-50	0.060	0.036
DINO ResNet-50	1.425	0.023
ViT	0.138	0.007
DINO ViT-S/8	0.047	0.042

These values are described as far better than Text-To-Concept and TEXTER, and especially better than Random [2605.16877]. The negative rate is particularly important because many baseline explanations correspond to directions that actually decrease the class logit; FaithTrace reduces that failure mode to near zero on several models [2605.16877].

The influence-curve results show the same pattern at finite perturbation scales. Under the top-3 setting, summed insertion and deletion scores are consistently much larger for FaithTrace [2605.16877]. The paper gives examples: for ResNet-18, insertion/deletion sums are (0.870/1.925) (in units of (10^{-1})) for FaithTrace, versus (0.159/0.933) for TEXTER; for DINO ResNet-50, FaithTrace gets (0.734/3.290), while baselines even have negative insertion sums [2605.16877].

The method is also evaluated on 100 misclassified images per model. Under top-3 on ResNet-18, FaithTrace obtains directional-score mean 0.190 with negative rate 0.000, versus 0.068 / 0.277 for TEXTER and 0.053 / 0.290 for Text-To-Concept [2605.16877]. Similar patterns hold across all models, often with zero or near-zero negative rate for FaithTrace [2605.16877]. This matters diagnostically because the explanations can reveal cues driving wrong predictions rather than only generic class semantics.

Qualitative examples reinforce that point. For a brambling, baseline methods focus on seeds or “feeding on ground,” whereas FaithTrace identifies “orange mottled plumage.” For a bee eater, baselines mention “distinctive chirping sound,” while FaithTrace gives “iridescent green plumage.” For a ring-necked snake, baselines are distracted by a human hand or generic snake properties; FaithTrace highlights the “yellow contrast” around the neck. For a goldfish, FaithTrace identifies “glass aquarium” as the actual driver, exposing contextual bias. In a misclassification where a jay is predicted as a little blue heron, FaithTrace outputs “wetland habitat,” aligning with the observation that many heron training images contain wetlands [2605.16877].

The paper further validates the semantics of text-induced directions by feature visualization using MACO. It synthesizes images whose classifier features match
[
z_{\text{target}} = z_f + \alpha \hat v_t(x),
]
with visualization defined by
[
x_{\text{vis}} = \underset{x' \in \mathcal{X}}{\operatorname{argmin}}\, \bigl|f(x') - z_{\text{target}}\bigr|_2^2.
]
With (\alpha=100), the visualizations enhance the corresponding visual cue, such as “sandy-colored plumage,” “vibrant blue plumage,” or “perched on rock” [2605.16877].

5. Relation to broader faithfulness-tracing research

The specific FaithTrace method belongs to a broader faithfulness agenda that also includes chain-of-thought tracing, process evaluation, executable evidence tracking, and provenance analysis. This broader interpretation is not a second formal definition of FaithTrace in the cited paper; rather, it is suggested by multiple adjacent works that explicitly discuss “FaithTrace”-style systems or projects [2510.22362] [2603.05167] [2605.24286] [2606.16151].

A central concern in that literature is that explanations may be plausible without reflecting the computation that produced the answer. “Mapping Faithful Reasoning in Language Models” introduces Concept Walk, an activation-space method that traces how a model’s internal representation of a target concept evolves across chain-of-thought steps under perturbation [2510.22362]. In that framework, easy cases show small and transient deviations in internal concept alignment after CoT perturbations, interpreted as CoT-as-rationalisation, whereas hard cases show sustained changes consistent with CoT-as-computation [2510.22362]. This suggests a temporal, internal-state analogue of FaithTrace’s shift from similarity to influence.

“Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning” formulates faithful CoT through three properties: sufficiency, completeness, and necessity [2605.24286]. The formal criteria are
[
H(A \mid C) \text{ is low},
\qquad
I(P; A \mid C) \approx 0,
\qquad
A \text{ depends causally on } C
]
[2605.24286]. That work introduces entropy-based, masked-KL, and gradient-based diagnostics and argues that low (H_q(A\mid C)), low Grad-DE, and high Grad-Nec are the most robust internal signals for FaithTrace-like auditing [2605.24286].

Benchmark work makes the point even more sharply. C(^2)-Faith decomposes process faithfulness into causality and coverage, showing that model rankings depend strongly on task framing, all judges exhibit a substantial gap between detecting an error and localizing it, and coverage judgments are systematically inflated for incomplete reasoning [2603.05167]. BonaFide goes further by meta-evaluating faithfulness metrics against task-derived ground truth and reporting that most metrics perform near chance, with the best metric reaching only 0.70 AUROC at the CoT level and another reaching 0.59 at the step level [2605.25052]. GRACE complements these by introducing a human-annotated step-level benchmark for context-grounded reasoning. On GRACE-test, 49.5% of traces with at least one unfaithful step still reach the correct final answer; the rate is 47.3% on the inference track and 51.6% on the grounding track [2606.16151].

A plausible implication is that FaithTrace, as a general research direction, is best understood not as one metric but as a family of procedures for verifying whether an explanatory object—text, concept direction, trajectory, provenance span, or executable trace—is functionally tied to the model behavior it is meant to explain. That interpretation is also supported by executable-trace systems such as ForestHG-Trace, which treats reasoning as deterministic operator execution over ecological hypergraphs, with
[
(\mathcal{S}{t+1}, r_t) = f{o_t}(\mathcal{S}t, \mathbf{u}_t)
]
and a replayable trace
[
\mathcal{T} = \left(\hat{a}, {\mathcal{S}_t}{t=0}^{T}, {a_t}_{t=0}^{T-1}\right)
]
[2605.27590].

6. Limitations, caveats, and research directions

The FaithTrace paper identifies three main limitations. The first is concept-bank coverage: because the method selects from a finite set (\mathcal{B}(x,c)), it can only explain using concepts present in that bank. If the candidate descriptions omit the true decision-driving cue, influence-based ranking cannot recover it [2605.16877]. The second is dependence on the quality of classifier–CLIP alignment; if the aligner poorly maps classifier features into a semantically meaningful language-connected space, the text-induced directions may be less reliable [2605.16877]. The third is that the influence measure is local and first-order: directional derivatives or moderate finite steps may not capture highly nonlinear behavior far from the original representation [2605.16877].

The method is also computationally heavier than pure retrieval by cosine similarity because each candidate text requires both a CLIP-similarity gradient and a classifier-logit directional derivative, though the paper describes it as still practical for the bank size used [2605.16877]. More broadly, the framework depends on external components—CLIP plus an LLM and a VLM—for candidate bank construction [2605.16877].

The broader faithfulness literature suggests additional cautions. Process-faithfulness numbers are highly sensitive to evaluator design: three classifiers applied to the same 10,276 influenced traces produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, with non-overlapping 95% confidence intervals [2603.20172]. BonaFide argues that existing faithfulness metrics often measure importance, semantic utility, plausibility, or output-level self-consistency rather than faithfulness itself [2605.25052]. GRACE shows that step-level faithfulness and final-answer correctness can diverge sharply [2606.16151]. These results caution against treating any single faithfulness score as objective across tasks, modalities, or explanation types.

Within explainable AI more broadly, related work such as DeepFaith proposes a unified optimal explanation objective over multiple faithfulness metrics and trains a learned explainer that generates explanations in a single forward pass without accessing the model being explained at deployment [2508.03586]. FaCT, by contrast, makes concepts part of the predictive computation itself and derives exact decompositions in which the output logit is the sum of concept contributions and concept activation is the sum of pixel contributions [2510.25512]. These neighboring approaches suggest two future directions for FaithTrace: stronger amortized explanation models and more mechanistically embedded explanation architectures.

The resulting landscape is not unified around one formal object. FaithTrace, in the narrow sense, is an influence-based textual explanation method for image classifiers [2605.16877]. In the broader sense suggested by adjacent work, it names a research program concerned with whether an explanation is actually used by the model, whether its evidence is recoverable, and whether the exposed trace is sufficiently faithful for oversight [2510.22362] [2605.24286] [2606.16151]. That broader program remains open because semantic plausibility, behavioral sensitivity, structural validity, step-level grounding, and mechanistic mediation are related but non-identical targets.