Study sets exact bounds on how model edits weaken OpenStamp watermarks in Llama-2-7B and Mistral-7B
INT8 quantization and 30% pruning keep detection at 85.6-94.8%, but 500-step LoRA fine-tuning cuts it to 28.8% and 11.6%, the authors report.
What happened
Some AI text carries a hidden statistical watermark that lets a checker tell which model wrote it. Open-weight models can be edited by anyone, for example by shrinking them or fine-tuning them. This paper measures how much those edits weaken the watermark signal.
Why it matters to you
If edits erase the watermark, it is harder to trace AI-generated text back to its source model. Companies relying on watermarks for provenance need to know which edits break them.
The takeaway
Light edits mostly leave the watermark detectable, but fine-tuning can largely erase it.
What it means for you
The residual-span bound gives a way to estimate how much span a modification needs to reach a target evidence level, instead of relying only on empirical attack tests.
Watermarking a released open-weight model offers weak protection against anyone who fine-tunes it with LoRA.