LEV starts from language models whose repeated application of the same layers was already learned during pretraining.
Looped decision models spend extra computation on harder questions
LEV repeatedly applies the same pretrained layers, improving fixed-choice accuracy while dynamically limiting computation on easier inputs.
Top University
Dingjie Song · Zhiling Yan · Yuxuan Zhang · Lichao Sun
Lehigh University · University of British Columbia · Vector Institute
Research Digest··3 min read
Song et al.
§