New method speeds up language model inference by half

Original: Pair-In, Pair-Out: Latent Multi-Token Prediction for Efficient LLMs

Source: arxiv.org ↗