12h ago

Continuous Trajectories Enable Over 5x Faster Non-Autoregressive Language Model Decoding

——0——
Original post
Luca Ambrogioni#1822@LUCAAMBOPGrigory SapunovGrigory Sapunov@CHE_SHR_CAT

1/ Non-autoregressive language models promised massive parallel speedups, but aggressive decoding always led to catastrophic quality collapse. Until now. By replacing rigid discrete token choices with soft continuous trajectories, we can now decode >5x faster. 🧵

4:30 PM · May 21, 2026 View on X
QUOTE POSTLuca Ambrogioni#1822Luca Ambrogioni@LUCAAMB

The way forward for discrete DLM is to turn them into continuous DLM ;)

Grigory SapunovGrigory Sapunov@che_shr_cat

1/ Non-autoregressive language models promised massive parallel speedups, but aggressive decoding always led to catastrophic quality collapse. Until now. By replacing rigid discrete token choices with soft continuous trajectories, we can now decode >5x faster. 🧵

11:30 PM · May 21, 2026 · 2.4K Views
9:57 AM · May 22, 2026 · 547 Views
Continuous Trajectories Enable Over 5x Faster Non-Autoregressive Language Model Decoding · Digg