ECHO recursively applies the backbone model’s final layer to predict successive future tokens.
Shared language-model layers make multi-token prediction cheaper and more direct
ECHO recursively reuses a model’s final layer to predict future tokens, reducing training overhead while supporting speculative decoding without a separately trained drafter.
Chinese Tech
Yuxuan Zhou · Junyuan Shang · Tianyu Wang · Chunrui Zheng · Zichang Liu · Wenbin Chang · +2 more
ERNIE Team, Baidu · Independent Researcher
Research Digest··2 min read
Zhou et al.
§