Microsoft & Cornell Paper: Free Pause Tokens Boost Prediction With ~1.14x Training-Only Overhead

dair_ai · x · 2026-09-06

dairai highlights a Microsoft/Cornell paper on pause tokens: they buy the model extra compute per next-token prediction, paying only with a sequence position. Free pause tokens carry the same compute in a parallel prediction stream over a weight-shared backbone. At inference they add nothing to context length, leave the KV cache unchanged, and cost essentially no latency; a 1B model gains 2-3 centinats on next-token prediction. The cost moves to training — roughly 1.14x overhead versus an optimized pretraining pipeline. A notable alternative if thinking-token context costs are unaffordable.

Original post →

More from Infra

Infra channel →