Reproduce “Athanor: An Agent-Native System for Autonomous Post-Training Research” 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
Yufa Zhou, Yichen Wang, Jiakun Fan, Ziqi Shu, Zhichao Liu, Anru R. Zhang, Neiwen Ling, Zijun Wei
Athanor supports four agents post-training a base language model through asynchronous training, frozen evaluation, and records linking results to their code and checkpoints. In a 104-hour campaign using public data without teacher-model calls, the final six-task means were 67.3 on development and 48.2 on held-out, versus 66.5 and 47.9 for the instruct model; confidence intervals include zero.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.