Training Time Comparison: Multi-Token vs. Next-Token Prediction

This table (S5) quantifies the training time overhead of multi-token prediction relative to next-token prediction, demonstrating its computational efficiency across different LLM sizes.