Authors: Chenglu Li, Gökhan Gülfidan, Yinqi Zhang-Kopf Source: Computers and Education: AI, Vol 11 — Open Access (CC BY 4.0)
Key Findings
Applied gradient-based LLM unlearning to 3 models pre-trained on ~3M Algebra I tutoring data points. Unlearned PII and harmful content while maintaining math task utility. PII output rate and harmfulness substantially decreased post-unlearning. Demonstrates a practical approach to making LLM-based math tutors more responsible and privacy-preserving.