List your company

Watermark Decoding on Llama2-7B Tokenizer 400k tokens (test)

1.44Decoding Latency (µs/token)HeRo¹ (3)
-18.1596114.1377246.435378.7323Jul 6, 2026

Evaluation results

MethodLinks
HeRo¹ (3)Batch Size=642026.071.44
HeRo² (2,2)Batch Size=642026.071.48
HeRo¹ (3)Batch Size=322026.071.57
HeRo² (2,2)Batch Size=322026.071.59
HeRo¹ (3)Batch Size=162026.071.66
HeRo² (2,2)Batch Size=162026.071.85
HeRo¹ (3)Batch Size=42026.072.43
HeRo² (2,2)Batch Size=42026.073.08
HeRo¹ (3)Batch Size=12026.077.75
HeRo² (2,2)Batch Size=12026.0710.34
BiMark (20)Batch Size=42026.07130.22
BiMark (20)Batch Size=12026.07130.43
BiMark (20)Batch Size=162026.07130.68
BiMark (20)Batch Size=322026.07131.63
BiMark (20)Batch Size=642026.07133.23
MPAC (2, 2.0)Batch Size=42026.07479.3
MPAC (2, 2.0)Batch Size=642026.07480.43
MPAC (2, 2.0)Batch Size=12026.07488.13
MPAC (2, 2.0)Batch Size=322026.07489.15
MPAC (2, 2.0)Batch Size=162026.07491.43