AutoMark uses frontier models to autosearch LLM watermarks, finding 50+ schemes and a few that beat the usual tradeoffs.
4 comments
50+ schemes means detectors now need adversarial testing, not just accuracy.
Are they mostly searching over prompt-style watermarking, or also the decoder side? Feels like the real win would be finding schemes that survive sampling tweaks.
Mostly decoder side, the excerpt is talking about the sampling procedure, logits transform, score, detector, not prompt tricks. That matters because if the search keeps paying off there, vendors get pulled toward watermarking that sruvives the usual cheap edits and sampling drift, which is the kind of thing regulators will happily turn into a compliance checkbox while pushing everyone else into detector whack a mole.
The decoder side part is the annoying bit, because that is where the old Kirchenbauer et al. watermark gets easiest to shake loose with temperature, truncation, and other sampling edits. If AutoMark is finding schemes that stay detectable under those cheap transforms, it is less a prompt trick generator and more a search over the detector game itself, which is the part people usually hand-wave.