A new technique can flag whether your data trained an AI model in a single pass.
Researchers built JUMP (Joint Uncertainty-Guided Mask Probing), a method for auditing whether specific records were used to fine-tune discrete diffusion language models, a newer architecture that predicts several masked words at once instead of one at a time. Unlike a prior approach called SAMA, which averages signals across many random masks and needs 32 model evaluations per sample, JUMP targets the positions the model is least confident about, masks them together, and scores the gap between the fine-tuned and original model's predictions using just three forward passes. Tested across six domains from the MIMIR benchmark, JUMP raised detection accuracy, measured as ROC-AUC, from 0.819 to 0.902 on LLaDA-8B-Base and from 0.851 to 0.942 on Dream-v0-Base-7B.
Companies increasingly fine-tune open-weight models on proprietary or sensitive data, and there has been no cheap way to verify what actually went into that process afterward. A faster, more accurate membership-inference test gives auditors and data owners a practical check rather than a research curiosity that needs dozens of costly queries per record.
Membership inference has mostly been a cat-and-mouse game on autoregressive models like GPT; this result suggests diffusion language models will face the same scrutiny as they catch on.