JUMP: Efficient Membership Inference on Fine-Tuned Diffusion Language Models
The episode discusses the paper "JUMP: Efficient Membership Inference on Fine-Tuned Diffusion Language Models" by Yeachan Jun and Albert No. JUMP proposes a single-pass method for membership inference in fine-tuned diffusion language models. It works by using a reference model to select low-confidence positions, then jointly masking and querying those tokens, significantly reducing computational cost compared to previous methods.