Rufus-Air: An Open LLM Post-Training Recipe

arXiv · Chia-Yuan Chang, Renyuan Cheng, Rui Feng, Xiaotian Han, Yuan He, Hongye Jin, Linwei Li, Shiyang Li, Fenglin Liu, Xin Liu, Priyanka Nigam, Haoyang Wen, Zhenghao Xu, Zhuocheng Xu, Bing Yin, Qingyu Yin, Chao Zhang, Rongzhi Zhang, Zhihan Zhang, Zixuan Zhang, Zixuan Zhang, Tuo Zhao · 2026-09-27

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Rufus-Air delivers an open, reproducible eight-stage post-training pipeline that pushes GLM-4.5-Air beyond its official release.

What it does

Rufus-Air outlines a serial pipeline of eight stages—SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF—applied to the 106B-parameter GLM-4.5-Air-Base model. The authors document data sources, reward designs, infrastructure choices, stage ordering, and stage-wise results, emphasizing that the entire recipe uses public data and open-source components without new human annotation or a private teacher model. Key findings include the importance of a strong SFT foundation, difficulty filtering to keep RL prompts productive, and reward reliability as a guiding principle for stage sequencing. The final model improves over the official GLM-4.5-Air post-trained release and is competitive with other open models of similar size.

Why it matters

Post-training is the last lever to extract performance from large language models, yet reproducibility and engineering details are often hidden. By openly sharing a full recipe, the work lowers the barrier for the community to replicate and extend high-performing LLMs without proprietary data. The staged approach also provides a practical blueprint for balancing hard, verifiable rewards with softer, judge-based signals.

How it applies

Teams can adopt the eight-stage pipeline as a checklist, reusing the documented data splits, reward functions, and infrastructure scripts to fine-tune their own base models. The emphasis on difficulty filtering and reward reliability helps avoid common pitfalls in RL-based fine-tuning. Because the recipe is open-source, practitioners can experiment with alternative stages or model sizes while retaining a proven baseline.

The short version

A recipe that lets you bake a stronger LLM from publicly available ingredients.

My take

I see Rufus-Air as the open-source playbook for the next wave of LLM engineering. The staged progression from SFT to RLHF mirrors how we naturally add skills to an agent, and the emphasis on difficulty filtering is a practical guardrail against runaway RL loss. By treating infrastructure as a core component, the recipe forces us to think holistically about the training stack.

How it connects

• Open post-training pipelines fuel the democratization of high-performing LLMs, echoing the broader open-AI movement.
• The staged approach aligns with curriculum learning trends in deep RL and LLM fine-tuning.
• Reward reliability as a sequencing principle anticipates future standards for safe, reproducible RLHF.

Bottom line

Follow the eight-stage Rufus-Air recipe and turn a public base model into a competitive, multi-skill agent.

Takeaways

• An eight-stage pipeline (SFT → RLHF) systematically builds capabilities from basic to advanced.
• Diverse, high-quality SFT creates a strong capability floor for later RL stages.
• Difficulty filtering keeps RL prompts within a productive learning range.
• Reward reliability serves as a practical principle for ordering training stages.
• Infrastructure and engineering choices are integral to the recipe, not merely implementation details.

Topics: LLM fine-tuning, RLHF, post-training, instruction following, open source