Reproduce “Logit Laundering: Evading Data-Use Auditing in Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ruihan Hu, Wei Luo, Yu-Ming Shang, Jiakai Wang, Chuxuan Zhang, Mei Li, Xi Zhang, Meikang Qiu

Logit Laundering suppresses watermark signals in outputs at inference without changing base-model parameters. It subtracts scaled logits from an auxiliary estimator trained on watermarked and public text, leaving the training corpus unchanged. For Llama-3.2-1B-Instruct, Radioactive’s average per-book p-value rises from 1.995×10⁻⁴ to 0.9948, while book-question accuracy is 72.8% with evasion versus 73.4% without. These results indicate evasion under tested audits, not proof of non-use.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.