Cross-Model Attribution Divergence

• Cross-model attribution divergence is defined as the difference between attribution vectors computed by the same explainer on two models, elucidating changes in feature importance.
• It leverages methods like occlusion, SHAP, and spectral analysis to compare attribution shifts and assess the impact of model updates.
• Empirical studies show it effectively distinguishes benign tweaks from substantive updates using metrics such as DCE, BAC, and JSD.

Searching arXiv for the cited papers to ground the article in current preprints. Cross-model attribution divergence denotes a family of methods that compare the feature-attribution structure of two models, or two versions of a model, in order to characterize how their decision logic differs. In the formulation introduced by "Delta-Audit: Explaining What Changes When Models Change" (Hemmat et al., 27 Aug 2025), if $A$ and $B$ are supervised classifiers and $E$ is any feature-attribution explainer returning per-feature scores $\phi_f(x)\in\mathbb R^d$ for input $x\in\mathbb R^d$, then the divergence is represented locally by $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$. In this view, positive coordinates indicate features that became more important from $A$ to $B$, and negative coordinates indicate features whose importance dropped. Related work extends the same comparative logic to heterogeneous model pairs, including XGBoost versus a LLM on clinical tabular prediction (Dasula et al., 17 Jun 2026), and to generative-model provenance, where divergence is measured through spectral routing behavior in diffusion denoisers rather than tabular feature scores (Meshram et al., 26 Jun 2026). Taken together, these studies establish cross-model attribution divergence as an audit paradigm for model updates, epistemic uncertainty diagnosis, and source attribution.

1. Formal definition and conceptual scope

The core object is a difference between attribution vectors computed under a common explanatory protocol. For models $A$ and $B$, and an explainer $B$0, Delta-Attribution defines

$B$1

This converts any single-model local explainer into a comparative instrument that answers “What changed in the decision logic when we swapped $B$2 for $B$3?” rather than “Why did $B$4 make its decision on $B$5?” (Hemmat et al., 27 Aug 2025). The construction is explicitly model-agnostic: traditional local explainers such as SHAP, Integrated Gradients, and occlusion/clamping can be repurposed by computing $B$6 and $B$7 with the same explainer and then differencing them.

Within this framework, the semantic interpretation is local and directional. If $B$8, feature $B$9 became more important for input $E$0 in the transition from $E$1 to $E$2; if $E$3, its importance decreased (Hemmat et al., 27 Aug 2025). This local difference map can then be aggregated across a test set to quantify whether a model update is merely cosmetic or reflects a substantive reallocation of reliance.

A second instantiation appears in clinical tabular prediction, where the compared objects are not two versions of the same architecture but XGBoost and Qwen 2.5 7B Instruct. There, cross-model attribution divergence is operationalized through an Attribution Disagreement Score (ADS) derived from rank disagreement between SHAP attributions for XGBoost and self-reported top-$E$4 feature attributions for the LLM (Dasula et al., 17 Jun 2026). A third instantiation generalizes the comparative idea beyond feature-importance vectors: Spectral Denoising Signatures (SDS) compare candidate diffusion models through how their denoisers redistribute energy across spatial frequency bands, and KL divergence between ring-wise coupling distributions is used to quantify cross-model attribution divergence in spectral routing behavior (Meshram et al., 26 Jun 2026).

This suggests that cross-model attribution divergence is best understood not as a single metric but as a comparative explanatory template. Its specific form depends on the attribution object being compared: per-feature tabular scores, ranked explanatory lists, or spectral coupling distributions.

2. Delta-Attribution as update auditing

In the update-audit setting, the objective is to explain what changes when models change. The paper introduces Delta-Attribution as a lightweight audit mechanism for model updates involving new hyperparameters, kernels, depths, solvers, or data (Hemmat et al., 27 Aug 2025). The empirical study covers 45 settings consisting of five classical families—Logistic Regression, SVC, Random Forests, Gradient Boosting, and $E$5NN—across three datasets, with three A/B pairs per family.

The method is instantiated via fast occlusion/clamping in standardized feature space, with a class-anchored margin and baseline averaging. The comparison uses the same baseline(s) and the same anchored class $E$6 for both models, ensuring a clean comparison (Hemmat et al., 27 Aug 2025). This shared explanatory protocol is essential: without it, observed differences could reflect explainer inconsistency rather than model divergence.

The practical significance of the method lies in its ability to separate benign from behaviorally meaningful changes. “Cosmetic” tweaks, such as SVC \texttt{gamma=scale} versus \texttt{auto} or changes in $E$7NN search procedure, yield rank-overlap@10$E$8 and DCE$E$9, indicating no meaningful reliance shift (Hemmat et al., 27 Aug 2025). By contrast, inductive-bias changes, including SVC poly$\phi_f(x)\in\mathbb R^d$0rbf on Breast Cancer and a Random Forest feature-rule swap on Digits, produce large, behaviour-aligned deltas, with BAC values near $\phi_f(x)\in\mathbb R^d$1 and nonzero DCE (Hemmat et al., 27 Aug 2025).

The paper’s conclusion frames this as an audit complement to accuracy: cross-model attribution divergence distinguishes benign parameter tweaks from structurally meaningful updates, flags risky reliance redistributions such as high magnitude with low BAC, and provides actionable thresholds for continuous integration of model updates (Hemmat et al., 27 Aug 2025). A plausible implication is that deployment decisions should not be based solely on aggregate performance deltas when model reliance may have shifted materially.

3. Metrics for quantifying divergence

The Delta-Attribution Quality Suite evaluates $\phi_f(x)\in\mathbb R^d$2 along four axes: magnitude/sparsity, agreement/shift, behavioural alignment, and robustness (Hemmat et al., 27 Aug 2025). Let $\phi_f(x)\in\mathbb R^d$3 and define the normalized divergence vector

$\phi_f(x)\in\mathbb R^d$4

skipping samples with $\phi_f(x)\in\mathbb R^d$5.

For magnitude and sparsity, the suite includes $\phi_f(x)\in\mathbb R^d$6-magnitude,

$\phi_f(x)\in\mathbb R^d$7

Top-$\phi_f(x)\in\mathbb R^d$8 concentration with $\phi_f(x)\in\mathbb R^d$9,

$x\in\mathbb R^d$0

and entropy,

$x\in\mathbb R^d$1

These distinguish large but concentrated shifts from large but diffuse ones (Hemmat et al., 27 Aug 2025).

For agreement and distributional shift, two measures are central. Rank-overlap@10 is the Jaccard overlap of the top-10 features by $x\in\mathbb R^d$2 and $x\in\mathbb R^d$3,

$x\in\mathbb R^d$4

Jensen–Shannon divergence is computed between the normalized absolute-attribution distributions

$x\in\mathbb R^d$5

with

$x\in\mathbb R^d$6

A high JSD indicates true redistribution of attribution mass rather than simple global rescaling (Hemmat et al., 27 Aug 2025).

For behavioural alignment, the framework defines $x\in\mathbb R^d$7 as the reference class and $x\in\mathbb R^d$8 as the change in model score on that class. Delta Conservation Error (DCE) measures how closely additive attribution differences sum to the actual score change,

$x\in\mathbb R^d$9

Because occlusion is not path-additive, DCE need not vanish; small DCE indicates faithful additive reasoning (Hemmat et al., 27 Aug 2025). Behaviour–Attribution Coupling (BAC) is the Pearson correlation between $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$0 and $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$1 across the test set. CO$\Delta\phi(x)=\phi_B(x)-\phi_A(x)$2F measures how much of $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$3’s mass falls on the top $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$4 features globally important for model $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$5, separately for samples where $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$6 fixes an error and where $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$7 regresses (Hemmat et al., 27 Aug 2025).

In the clinical LLM study, a distinct divergence metric is used. Attribution vectors are converted into ranked top-$\Delta\phi(x)=\phi_B(x)-\phi_A(x)$8 lists, and Kendall’s $\Delta\phi(x)=\phi_B(x)-\phi_A(x)$9 between the XGBoost and LLM rankings is transformed into ADS:

$A$0

ADS ranges from $A$1 for perfect agreement to $A$2 for perfect anti-correlation, with $A$3 denoting zero correlation (Dasula et al., 17 Jun 2026). Secondary metrics are Jaccard overlap, top-1 match rate, and directional agreement on shared features.

In the diffusion-model setting, divergence is measured at the level of spectral coupling distributions. After constructing row-normalized coupling matrices $A$4, KL divergence between ring-wise distributions,

$A$5

quantifies per-ring spectral divergence; summed or averaged over $A$6, it measures cross-model attribution divergence in spectral routing behavior (Meshram et al., 26 Jun 2026).

4. Algorithmic instantiations

The update-audit instantiation in Delta-Audit follows a five-step pipeline (Hemmat et al., 27 Aug 2025). First, a shared StandardScaler is trained on model $A$7’s training data and applied to both $A$8 and $A$9, placing inputs in a common standardized space. Second, for each test sample $B$0, the reference class is anchored as $B$1, and the scalar score $B$2 is defined as the margin or log-odds of class $B$3. Third, two baselines, $B$4 and $B$5, are computed from the training set in standardized space and averaged:

$B$6

Fourth, fast occlusion/clamping computes per-feature attributions for each model by replacing coordinate $B$7 with $B$8 and evaluating

$B$9

This requires $A$0 forward passes per sample per model. Fifth, $A$1 is computed and scored with the Quality Suite (Hemmat et al., 27 Aug 2025).

The clinical tabular setting uses a heterogeneous attribution pipeline. For XGBoost, TreeSHAP computes exact Shapley values $A$2, yielding $A$3 (Dasula et al., 17 Jun 2026). For the LLM, the model is prompted to return its top-$A$4 features and associated importance scores in JSON; these are embedded into a $A$5-dimensional vector by zero-padding unlisted features, producing $A$6 (Dasula et al., 17 Jun 2026). The vectors are then converted into top-$A$7 rankings for Kendall-$A$8 computation and ADS evaluation.

The diffusion-model provenance setting is algorithmically different because the attribution object is the denoiser’s spectral response. SDS perturbs latent variables with band-limited Fourier noise, injects the perturbation at a chosen timestep, computes the denoiser residual

$A$9

and measures how residual energy is redistributed across spatial-frequency rings (Meshram et al., 26 Jun 2026). Repeating the procedure over scales, timesteps, source rings, and random draws yields a tensor

$B$0

which is flattened into a signature vector $B$1. A model prototype is then the centroid

$B$2

Attribution reduces to comparing a query signature to stored prototypes with cosine distance, $B$3 distance, Mahalanobis distance, or a linear classifier (Meshram et al., 26 Jun 2026).

These three instantiations share a common structure: compute model-specific explanatory objects under a controlled protocol, convert them into a comparable representation, and quantify divergence with metrics that distinguish simple scaling changes from substantive shifts in reliance or routing.

5. Empirical patterns across domains

The most explicit empirical taxonomy appears in Delta-Audit. Across 45 A/B pairs, cross-model attribution divergence cleanly separates three regimes (Hemmat et al., 27 Aug 2025). Cosmetic tweaks produce tiny $B$4, approximately $B$5–$B$6, with RankOverlap@10 approximately $B$7, DCE approximately $B$8, and BAC approximately $B$9 or undefined. Inductive-bias shifts produce large $B$00, for example $B$01–$B$02, high BAC of approximately $B$03–$B$04, moderate DCE of $B$05–$B$06, and JSD up to approximately $B$07. $B$08NN parameter changes yield large but diffuse $B$09 of approximately $B$10–$B$11, low BAC of approximately $B$12–$B$13, and moderate DCE of $B$14–$B$15, flagging reliance shifts not tightly coupled to outcome changes (Hemmat et al., 27 Aug 2025).

Representative quantitative results illustrate these cases. For breast_cancer–svc–pair3 (poly$B$16rbf), BAC is approximately $B$17, DCE approximately $B$18, and JSD approximately $B$19. For digits–rf–pair3 (max_feat:$B$20), BAC is approximately $B$21 and DCE approximately $B$22. For breast_cancer–gb–pair3 (depth $B$23), JSD is approximately $B$24, the largest reported redistribution. For wine–rf–pair2, Mag is approximately $B$25, BAC approximately $B$26, DCE approximately $B$27, and RankOverlap@10 approximately $B$28 (Hemmat et al., 27 Aug 2025).

In the clinical LLM setting, the results are organized around four findings (Dasula et al., 17 Jun 2026). The first is confidence invariance: zero-shot LLM confidence is approximately $B$29 whether accuracy is $B$30 or $B$31, and few-shot LLM confidence is approximately $B$32 whether accuracy is $B$33 or $B$34; verbalized confidence carries zero information, with AUROC approximately $B$35. The second is an inverse difficulty effect: when XGBoost confidence exceeds $B$36, XGBoost accuracy is $B$37 while the few-shot LLM is $B$38, but when XGBoost confidence lies in $B$39, XGBoost and LLM accuracies are $B$40 and $B$41, respectively (Dasula et al., 17 Jun 2026).

The third clinical finding is super-additive attribution alignment. On a 300-example test subset, zero-shot yields accuracy $B$42 and ADS $B$43; zero-shot plus SHAP injection yields $B$44 and $B$45; few-shot yields $B$46 and $B$47; few-shot plus SHAP yields $B$48 and $B$49 (Dasula et al., 17 Jun 2026). The combined reduction in ADS, $B$50, exceeds the sum of the individual reductions from zero-shot to zero-shot plus SHAP and from zero-shot to few-shot, and the accuracy gain from zero-shot to few-shot plus SHAP is likewise described as super-additive (Dasula et al., 17 Jun 2026). The fourth finding concerns calibration: a patient-level classifier built from ADS, Jaccard, directional agreement, top-1 match, XGB confidence, LLM confidence, and the absolute confidence gap produces markedly better reliability estimates than raw verbalized confidence, with the best calibrator reducing expected calibration error from $B$51 to $B$52 (Dasula et al., 17 Jun 2026).

In the diffusion setting, SDS achieves approximately $B$53 accuracy across eight diverse diffusion models and $B$54 under cross-domain prompt shift (Meshram et al., 26 Jun 2026). In the closed-set attribution experiment with 50 prototypes per model and 500 test images per model, LinearSVC on raw signatures yields $B$55 on SD-Prompts and $B$56 on MS-COCO, while zero-training argmin with cosine distance yields $B$57 and $B$58 (Meshram et al., 26 Jun 2026). On five models sharing the same VAE, reconstruction-based non-invasive baselines collapse to $B$59–$B$60, whereas SDS maintains approximately $B$61 across all eight models (Meshram et al., 26 Jun 2026).

6. Interpretation, limitations, and related controversies

A recurrent interpretive distinction is between divergence that is behaviorally aligned and divergence that is merely large. In Delta-Audit, high magnitude accompanied by high BAC indicates that attribution shifts track actual score changes, whereas high magnitude with low BAC flags reliance shifts that are not tightly coupled to outcome changes (Hemmat et al., 27 Aug 2025). This does not by itself prove risk, but the paper explicitly treats such cases as requiring deeper investigation or regularization. A plausible implication is that cross-model attribution divergence can function as a triage signal in model-governance workflows.

A second issue concerns faithfulness. The clinical study uses self-reported LLM attributions obtained from JSON-formatted top-$B$62 feature lists (Dasula et al., 17 Jun 2026). The paper identifies this as a limitation, noting that such attributions raise faithfulness questions. The practical argument is therefore narrower than in the update-audit setting: high ADS is used diagnostically as a proxy for epistemic unreliability, and the observed reductions in ADS under few-shot prompting and SHAP injection are interpreted together with improved accuracy and calibration rather than as stand-alone evidence of causal reasoning (Dasula et al., 17 Jun 2026).

A third issue is comparability across architectures and modalities. The diffusion-model work demonstrates that the same comparative principle can be transferred from tabular feature attributions to spectral coupling signatures in denoiser Jacobians (Meshram et al., 26 Jun 2026). This suggests that “attribution” in cross-model divergence need not be restricted to feature-importance vectors, provided the compared objects have a stable interpretive role as model-specific explanatory signatures. At the same time, the SDS setting imposes stronger access assumptions: it requires forward-pass access to the exact denoiser weights and assumes a closed set in which the queried model appears in the registry (Meshram et al., 26 Jun 2026).

The three papers also delineate distinct operational limits. Delta-Audit is framed for supervised classifiers and model updates (Hemmat et al., 27 Aug 2025). The clinical study is limited to a single LLM, a single AKI task on MIMIC-IV with $B$63, and non-exhaustive prompt strategies (Dasula et al., 17 Jun 2026). SDS does not address open-set attribution, black-box APIs, or tamper detection for post-generation edits (Meshram et al., 26 Jun 2026). These are not contradictions so much as boundary conditions on where current formulations of cross-model attribution divergence are well specified.

7. Research directions and broader significance

Several future directions are stated explicitly. In the clinical setting, proposed lines of work include RL-based confidence internalization using the calibrator’s predicted probability of correctness as a reward, multi-task calibration across structured prediction tasks, benchmarking ADS against semantic-entropy, token-entropy, and conformal prediction, and clinician-in-the-loop validation of SHAP injections and calibrator outputs (Dasula et al., 17 Jun 2026). These proposals frame cross-model attribution divergence not only as a diagnostic but also as a supervisory signal for improving epistemic calibration.

In the diffusion setting, extensions include open-set detection via outlier thresholds or one-class classifiers on signature space, adaptation to other generative families such as GANs and autoregressive models, fine-grained analysis of fine-tuning, quantization, or distillation artifacts, and adversarial defenses against attempts to evict or mimic spectral fingerprints (Meshram et al., 26 Jun 2026). This suggests an emerging role for attribution divergence in provenance verification and intellectual-property enforcement, especially where output-space evidence is unreliable.

In the update-audit setting, the main significance lies in integrating explanatory comparison into continuous model evaluation. Delta-Attribution is presented as a lightweight update audit that complements accuracy by distinguishing benign changes from behaviorally meaningful or risky reliance shifts (Hemmat et al., 27 Aug 2025). The broader implication is methodological: when models are updated, compared, or combined, explanatory divergence can provide an additional axis of evaluation that is orthogonal to standard predictive metrics.

Cross-model attribution divergence therefore occupies an intermediate position between explainability, uncertainty quantification, and model forensics. In tabular supervised learning, it exposes what changed between model versions; in hybrid clinical systems, it diagnoses where an LLM “doesn’t know what it doesn’t know”; and in diffusion provenance, it identifies source models through intrinsic denoising behavior (Hemmat et al., 27 Aug 2025, Dasula et al., 17 Jun 2026, Meshram et al., 26 Jun 2026). Across these domains, the unifying premise is that differences in explanatory structure can reveal model distinctions that accuracy, confidence, or output inspection alone may leave opaque.