John Kirchenbauer

Multi-Token Prediction via Self-Distillation

Researchers have developed a novel self-distillation technique to accelerate language model inference. This method transforms a standard autoregressive model into a faster multi-token predictor without needing auxiliary…