OpenMark Results

OpenMark

OpenMark is our method in OpenMarkX and its default. It learns a mark that is precise to detect and hard to edit away.

PreciseThe watermark rule is learned together with the model, so the model learns it closely.
Watermark rule
Model
RobustDuring training, OpenMark simulates the edit that would erase the mark fastest, and trains the model so that edit barely hurts it.
Model
Worst-case editerases the mark fastest

Results

Alice releases her model watermarked with OpenMark, and users edit it. How much of the mark is left after each edit, and how good is the model?

Alice’s LLMas released
Detection≥99%traced to her
Fidelity after watermarking4.1perplexity; unmarked model 3.5
Quantized4-bit
≥99%
Pruned50% of weights
≥99%
Mergedwith another model, t = 0.3
≥99%
Fine-tuned300 steps on math
61.5%
+ your editone file, like a method

After release, users edit the model and publish their own copies. Each branch is one edit of Alice’s model; the number is how many of its texts are still traced to her (OpenMark on Granite-4.2-3B, 200 generated texts per model, 1% false positives).

m2 = omx.modify(m, "merge:partner=PARTNER,ratio=0.3")Every edit plugs in the same way.

Compared with other watermarking methods (LLaMA-3-8B)

Detection after each edit · Llama-3-8B · texts flagged at 1% false positives

OpenMark keeps the most mark under merging and fine-tuning, the edits that erase marks fastest.