Work

Explainability-Based Token Replacement on LLM-Generated Text

Year: 2025

Type: preprint

Abstract: Code, contestant models, and analysis for the explainability-based token-replacement experiments on AI-generated-text detectors. SHAP and LIME identify the tokens most influential to detector decisions; those tokens are then systematically replaced on the CLIN33 and AuTexTification corpora to expose adversarial weaknesses and pathways to greater robustness. Archived snapshot for the v1.0-thesis release accompanying the doctoral thesis Let Me Explain! Explainable NLP for Understanding Large Language Models (Utrecht University, 2026).

Source: arXiv (Cornell University)

DOI: 10.48550/arxiv.2506.04050

Authors: Mohammadi, Hadi, Giachanou, Anastasia, Oberski, Daniel L., Bagheri, Ayoub

Institution: Utrecht University

Language: English

Topic: Explainable Artificial Intelligence (XAI)

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funder: Nederlandse Organisatie voor Wetenschappelijk Onderzoek

Awards: G492894576, Advancing social science with valid measures derived from incidental data

OpenAlex ID: W4416074985

Cites:

Cited by:

Related to: