Reference-Based Distillation Detection in LLMs

Rajat Rawat, Sizhe Chen, Akshay Anand +3

Model distillation -- training on outputs from stronger third-party models -- is widely used to boost performance, but raises concerns about unfair advantages and policy violations…