Skip to content
InfoResearchPreprintLLM-specific

Watermark Evidence under Open-Weight LLM Modification: Exact Bounds for Low-Rank Output Watermarks

Published
Record updated
View JSON

Summary

This paper studies how modifying open-weight language models changes the evidence available to an OpenStamp-style watermark verifier. The authors derive exact bounds on expected log-likelihood-ratio evidence under a residual-span constraint, plus a high-confidence bound on miss probability for sequence-level detection. In experiments on Llama-2-7B and Mistral-7B-v0.3, INT8 quantization and 30% pruning kept test true-positive rates at 85.6-94.8%, while 500-step LoRA fine-tuning reduced them to 28.8% and 11.6%.