SPECULATIVE DECODING FOR RL TRAINING

Summary from the TLDR AI newsletter

Speculative decoding was applied to RL rollouts without changing output distributions, delivering up to 1.8x throughput gains and projected 2.5x end-to-end speedups at scale.