MoE routing doesn't kill speculative decoding, shared experts preserve speedups

MoE routing isn't killing speculative decoding speedups the way people assumed. The usual fear is that activating different experts per token blows up latency, but this report says the draft-and-verify path actually benefits instead of getting canceled out.

The key is reuse: adjacent tokens share about 38% of experts, around 3-6× higher than an independence baseline, and that held across 13 tasks and 7 languages. That overlap keeps routing costs from exploding and lets speculation stack gains. It also finds a non-monotonic speed curve with a sweet spot at moderate batch sizes, plus a batch-size-1 bonus from amortizing fixed overhead.