1. 2

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

Paper proposes replaying expert RL training logs to merge experts into one model.

  • Method: replays logged rollouts through the student in training order.
  • Cost: no new sampling, about a seventh of expert training compute.
  • Results: claims to beat online RL and MOPD (multi-teacher distillation).