Reproduce “Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate” 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
Haocheng Ye, Aoting Hu, Xinwei Zhang, Xunzhu Tang, Shuchao Pang, Jason Xue
The authors evaluate decoding-time watermarking in controlled retrieval-augmented generation using paired answers with the same evidence and decoding conditions. They measure net factual changes and trace part of the effect to reduced attention to supporting context. Protecting factual tokens and compensating attention lowers KGW’s net accuracy penalty from 12.9% to 0.9%, but residual errors remain in this controlled setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.