Reproduce “Recursive Offline Distillation via Entropy Gain Sampling and Answer Mixing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andrey Goncharov, Daniil Vyazhev, Simon Epanov, Alexey Zaytsev
The authors present recursive offline distillation, which selects pre-generated reasoning traces by comparing a student’s first-answer-token uncertainty with that of a larger proxy, then updates the selection each training epoch. The method also trains on direct answers. Across three small models, it reaches mean last-epoch MMLU-Pro accuracy of 57.15 ± 1.01, versus 51.61 for direct-trace distillation, in this benchmark setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.