ProToMEx arXiv:2609.04265 Reports 30-40x Speedup Over SHAP and LIME on Tabular Benchmarks

The framework replaces per-instance feature attribution with a model-agnostic Probabilistic Topic Model that extracts latent topics representing distinct classification rationales. Global topics describe overall model behaviour; local inference decomposes individual predictions into mixtures of these topics. Training occurs once; subsequent explanations require only topic inference rather than repeated perturbation sampling. Empirical results cite direct wall-clock measurements across multiple tabular datasets, showing the 30-40x reduction in per-explanation time alongside fidelity scores statistically indistinguishable from SHAP and LIME baselines. The paper records no degradation on synthetic data where ground-truth combinatorial patterns are known. Operationally this shifts explainability from offline auditing to online serving constraints, where sub-second latency budgets previously excluded post-hoc methods. The topic representation also surfaces co-occurring reasons that feature-attribution vectors obscure. No production deployments or follow-up benchmarks appear in the submission. Subsequent work would need to validate topic stability across retraining cycles and extend the approach beyond tabular inputs.