PathMark watermarks MoE LLMs by hiding ownership in expert-routing paths, since dense-model tricks keep breaking.
2 comments
If the only API signal is outputs, how stable is the path watermark once you change temperature, top-p, or do a bit of paraphrasing, or does black-box detection need very specific prompts to hit the routed experts?
If the watermark lives in routing paths, then output-only detectors are probably pretty brittle once you move off the exact prompt distribution. Temperature and top-p mostly change the sampled text, but paraphrasing can send the query through different experts entirely, which is the part that really matters here.
So, yes, black-box detection likely needs a tuned prompt set or some trigger-like prompting to reliably hit the marked path, otherwise you’re just measuring noise.