ATT consistency

ATT consistency asks whether the same input, model, explanation configuration, and intended operating conditions produce reproducible attributions across repeated executions.

Unlike ATT stability, the evaluated input is not deliberately changed. Consistency isolates execution variability arising from random perturbation sampling, surrogate fitting, stochastic model generation, provider behaviour, hardware, or other uncontrolled factors.

General protocol

For a fixed input (x), run the explanation procedure (R) times:

a¹(x), a²(x), ..., aᴿ(x)

Compare the attribution vectors, rankings, selected feature sets, and any relevant output signals across runs.

Sources of inconsistency

• random perturbation sampling;
• random initialisation or optimisation in the surrogate;
• non-deterministic black-box inference;
• LLM sampling temperature and provider-side variability;
• stochastic segmentation or clustering;
• changes in retrieved evidence for RAG systems;
• unstable feature construction or alignment;
• different distance, kernel, or hyperparameter choices when these are intended to be fixed.

Modality-specific interpretation

Modality	Repeated object	Consistency evidence
Image classification	Same image and explanation configuration	Similar superpixel rankings and attribution magnitudes
LLM / gSMILE	Same prompt, model, parameters, and provider setup	Similar token weights and generated response behaviour
Image editing	Same source image, instruction, seed policy, and editor	Similar instruction attributions and edited-image effects
Point cloud	Same point cloud and preprocessing	Similar point/cluster importance despite execution randomness
KG-RAG	Same query, graph snapshot, retrieval configuration, and model	Similar evidence ranking and attribution paths
ConceptSMILE	Same image, concept pathway, prompt, and perturbation design	Low variation in concept scores and feature importance

Measures

ConceptSMILE treats lower variance and standard deviation across repeated executions as stronger reproducibility. Depending on the modality, report:

• per-feature mean, variance, and standard deviation;
• confidence intervals for attribution values;
• Jaccard overlap of top-k feature sets;
• rank correlation between repeated attribution orderings;
• frequency with which each feature enters the top-k set;
• variance of summary metrics such as fidelity or coverage.

A simple user-side analysis can use XWhy result coefficients:

import numpy as np

runs = [explainer.explain(instance) for _ in range(10)]
coefficient_matrix = np.vstack([result.coefficients for result in runs])

mean_attribution = coefficient_matrix.mean(axis=0)
std_attribution = coefficient_matrix.std(axis=0)
variance = coefficient_matrix.var(axis=0)

This assumes that the feature space is identical across runs. For image explanations, regenerated superpixels may differ; for point-cloud explanations, clusters may change; and for LLMs, word or token segmentation must remain aligned.

Current XWhy support

XWhy exposes the attribution vector and raw evaluation data needed for repeated-run analysis, but there is currently no dedicated public ATTConsistency function.

Relevant controls include:

• explainer random seeds;
• LLM temperature and provider options;
• perturbation count;
• surrogate type and automatic surrogate selection;
• distance metric and locality weighting;
• segmentation or feature-construction configuration.

For LLMs, setting temperature=0.0 and a fixed seed reduces known sampling variation, but an external provider may still be non-deterministic. Report the provider, model version, request parameters, date, and repeated outputs.

Consistency across hyperparameters

The gSMILE discussion also considers consistency across model runs or hyperparameter settings. This should be reported separately from strict repeated-run consistency:

• repeatability: same input and same configuration;
• configuration robustness: same input under deliberately varied but plausible settings.

Mixing these experiments into one score obscures whether variability came from randomness or an intentional design change.

Reporting checklist

Report:

• number of repeated runs;
• all fixed and varying parameters;
• random seeds and deterministic settings;
• provider/model version and retrieval snapshot where relevant;
• feature-alignment procedure;
• per-feature variability and aggregate overlap;
• uncertainty intervals rather than only one average score.

Consistency is not stability or validity

Repeatable explanations may still be inaccurate or unfaithful. Conversely, a stochastic model may produce some attribution variability even when the explanation method is behaving appropriately. Interpret consistency together with the task and model's expected randomness.

Research basis

• gSMILE defines ATT consistency through repeatable token weights and outputs for repeated use of the same prompt.
• ConceptSMILE evaluates reproducibility through repeated execution and reports variance and standard deviation as consistency indicators.