Reproduce “Rufus-Air: An Open LLM Post-Training Recipe” 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

Chia-Yuan Chang, Renyuan Cheng, Rui Feng, Xiaotian Han, Yuan He, Hongye Jin, Linwei Li, Shiyang Li and 14 more

The authors present Rufus-Air, a reproducible eight-stage post-training recipe for GLM-4.5-Air-Base, using open data and components without new human annotation or an in-house teacher. The pipeline progresses from SFT through verifiable-reward and agent training to RLHF. It scores 76.9 on prompt-strict IFBench versus 33.6 for the same-base official release, but is not shown optimal and trails it on Creative Writing.

Key insights

  • Reward reliability motivates the training order: deterministic mathematical and programming checks are used early, while more subjective search and preference rewards are delayed until the final stages.
  • SFT is treated as a capability-building foundation rather than a brief warm-up; its data mixture emphasizes broad coverage, valid multi-turn tool traces, long-context behavior, and decontamination.
  • The results highlight data curation and systems engineering as central parts of reproducible RL, but do not demonstrate that the selected stage order is globally optimal because comprehensive ablations are limited.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.