When New Generators Arrive: Lifelong Machine-Generated Text Attribution via Ridge Feature Transfer
Researchers have developed a new framework called RidgeFT to address the challenge of machine-generated text attribution, particularly when new language models are continuously introduced. This method allows attribution models to adapt to new generators without forgetting previously learned ones, a common issue with existing approaches. RidgeFT employs a lightweight, replay-free update mechanism that stores compact statistics for each generator and uses closed-form ridge regression for updates, outperforming baseline methods in multi-topic evaluations. AI
IMPACT Enhances the ability to track and attribute machine-generated text, crucial for accountability and misuse investigations as new models emerge.