TLDR Tech / AISpeculative Decoding for RL Training2026-05-01Speculative 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.Read the full story