When New Generators Arrive: Lifelong Machine-Generated Text Attribution via Ridge Feature Transfer

In three linesRidgeFT, a lifelong learning framework without exemplar replay, attributes machine-generated text to specific LLM sources. The method freezes an initially trained encoder, stores class-wise sufficient statistics, and uses closed-form ridge regression to adapt to new generators while retaining recognition of previous ones.

Summary generated by Claude — human-verified