CopyLoop accelerates looped transformer decoding using context copying
October 30, 2026
CopyLoop is a training-free algorithm that accelerates looped transformer decoding by proposing multi-token suffixes from repeated context. It uses shallow screening and continuous recurrent scheduling to filter incorrect branches early, reducing the computational cost of recurrent depth.
HOW THIS AFFECTS YOU
●
builderYou can achieve faster generation speeds for looped transformer models without retraining.
●
researcherThe shallow screening approach provides a new way to manage speculative decoding in recurrent architectures.