Generation of an output token sequence from an input token sequence using two language model neural networks

Tianlin Liu, Shangmin Guo, Daniele Calandriello, Quentin Berthet, Felipe Llinares López, Jessica Hoffmann, Lucas Dixon, Michal Valko, Mathieu Blondel

2025 · International Patent WO2025166256A1 (2025)

Abstract

A method of using a first and a second language model neural network to generate an output token sequence from an input token sequence, where at least the second network has been fine-tuned. The method combines scores from both networks based on a realignment parameter to enable control over the degree of alignment at decoding time without requiring retraining.

PDF · WO2025166256A1