Zinan Sheng

Archer: Adaptive Reuse of Cached Hidden States for Efficient Rollback in Diffusion Language Models

Adaptive Reuse of Cached Hidden States for Efficient Rollback (Archer) is introduced, a training-free KV caching method for rollback-capable DLMs that characterizes prompt reuse as a reversibility-aligned cache boundary, bounds its state-dependent approximation error, and gives a decoder-margin condition for preserving full-refresh decisions.