works on
candidate-aware decoding 1diffusion language models 1early exit decoding 1speedup 1zero-shot evaluation 1

From the 1 of 3 linked papers with an AI index.

collaborators
Showing cs.CLShow all

1 paper · 1 filter

cs.CL2026

Commit Locally, Exit Globally: Coordinating Adaptive Sampling and Early Exit in Diffusion Language Models

Chia-Ming Lee, Ming-Ching Chang, Shao-Kai Liu +3

The paper introduces LATCH, a training‑free, candidate‑aware early‑exit framework for diffusion language models that decides when to stop generation and where to accelerate decodin…