What question did this study set out to answer?

The aim is to enhance the fidelity of explanations in transformer models by optimizing feature corruption sequences.

March 4, 2026AlgorithmsOpen Access

EvoDropX:Evolutionary Optimization of Feature Corruption Sequences for Faithful Explanations of Transformer Models

View Full Paper

Ask AI

Bookmark

Authors

Demonstrates a new method to enhance explanation fidelity in transformer models, suggesting improved transparency in decision-making systems.

The aim is to enhance the fidelity of explanations in transformer models by optimizing feature corruption sequences.
Introduced EvoDropX as an optimization problem for explanation generation.
Leveraged Grammatical Evolution to evolve sequences of feature corruption.
Evaluated performance using the Symmetric Relevance Gain metric across multiple datasets and transformer models.
Compared EvoDropX to state-of-the-art explainability methods like SHAP and LIME.
EvoDropX significantly outperformed traditional xAI methods in explanation fidelity.
Achieved a 74.77% improvement in Symmetric Relevance Gain on the IMDB dataset with the BERT model.
Demonstrated consistent improvements across all tested dataset-model combinations.
Captured sentiment-bearing terms and maintained structural relationships in explanations.

Singh et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark

March 4, 2026AlgorithmsOpen Access

View Full Paper

Ask AI

Bookmark

Authors

Demonstrates a new method to enhance explanation fidelity in transformer models, suggesting improved transparency in decision-making systems.

The aim is to enhance the fidelity of explanations in transformer models by optimizing feature corruption sequences.
Introduced EvoDropX as an optimization problem for explanation generation.
Leveraged Grammatical Evolution to evolve sequences of feature corruption.
Evaluated performance using the Symmetric Relevance Gain metric across multiple datasets and transformer models.
Compared EvoDropX to state-of-the-art explainability methods like SHAP and LIME.
EvoDropX significantly outperformed traditional xAI methods in explanation fidelity.
Achieved a 74.77% improvement in Symmetric Relevance Gain on the IMDB dataset with the BERT model.
Demonstrated consistent improvements across all tested dataset-model combinations.
Captured sentiment-bearing terms and maintained structural relationships in explanations.

Singh et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark