notesum.ai

Published at November 29

Ensemble Watermarks for Large Language Models

cs.CL

Released Date: November 29, 2024

Authors: Georg Niess¹, Roman Kern²

Aff.: ¹Graz University of Technology; ²Know-Center GmbH & Graz University of Technology

Arxiv: http://arxiv.org/pdf/2411.19563v1

Refer to caption

Configuration	Human	Senso	Acro	Red-Green	Senso + Acro	Senso + Red-Green	Acro + Red-Green	All Three
Llama 3.1 8B (Strong)	0.34	80.41	28.52	49.14	89.35	84.19	55.67	95.19
Llama 3.2 3B (Strong)	0.97	85.11	31.39	54.05	90.61	90.29	64.08	95.79
Llama 3.1 8B (Medium)	2.42	58.06	28.23	34.14	73.92	70.70	46.77	82.53
Mistral 7B (Medium)	1.44	54.87	42.60	23.47	69.31	58.84	40.07	73.65
Llama 3.2 3B (Weak)	1.32	26.07	27.72	29.04	38.61	43.56	35.97	44.88