Amazon's AI Research division has open-sourced 'Rufus-Air: An Open LLM Post-Training Recipe' (arXiv: 2609.29421), an entirely reproducible eight-stage post-training pipeline built atop GLM-4.5-Air-Base (a 106B total / 12B active parameter Mixture-of-Experts model). Dispensing with proprietary teacher distillation and private human annotation, Rufus-Air cascades through eight distinct stages—SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF—leveraging solely public data and open tooling to beat the official GLM-4.5-Air checkpoint across standard benchmarks.

Key Takeaways

  • ✓Delivers the first fully reproducible 8-stage post-training recipe for 100B+ MoE architectures (106B-A12B GLM-4.5-Air).
  • ✓Relies entirely on publicly available datasets and verifiable programmatic reward models without proprietary teacher distillation.
  • ✓Validates the 'Reward Reliability Ordering' heuristic, sequencing stages from deterministic unit tests to soft LLM judge signals.
  • ✓Outperforms the official GLM-4.5-Air checkpoint across core evaluation suites, registering a 21.8% boost on multi-turn coding agent tasks.
  • ✓Full stage ordering configs, reward architectures, and distributed training recipes published openly on arXiv and Hugging Face.
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
🌐
ArXiv 论文预印本 (2609.29421)
arxiv.org/abs/2609.29421
↗
📝
Hugging Face Daily Papers 讨论主页
huggingface.co/papers/2609.29421
↗
📝
Amazon Science 官方研究主页
www.amazon.science
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points Post-training reinforcement learning has emerged as the definitive catalyst for advancing LLM reasoning and agentic problem solving. However, frontier post-training recipes remain proprietary trade secrets, obfuscating stage sequencing, reward modeling, and curriculum curation behind closed doors. Community reproduction attempts routinely falter due to reward hacking and catastrophic forgetting across multi-objective rollouts. ### Architecture Highlights & Internals Amazon AI structures Rufus-Air into a deterministic eight-stage sequential pipeline on GLM-4.5-Air-Base (106B-A12B MoE): SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. The architecture operates on the principle of Reward Reliability Ordering—progressing from strictly verifiable deterministic unit-test signals toward nuanced semantic judges. Rigorous difficulty filtering confines prompts to optimal learning gradients, while dedicated agent stages isolate multi-turn tool interaction policies. ### Authoritative Benchmarks & Measured Scores Rufus-Air decisively surpasses the official GLM-4.5-Air release: SWE-bench Verified pass rates climbed by 14.2%, while Tau-Bench and AgentBench success rates jumped by 21.8%. On AIME and MATH-500 competitions, mathematical deduction reached 52.4% and 84.6% respectively, matching the upper echelon of active 12B-class architectures. ### Developer Hands-on Guide Machine learning engineers can review the detailed distributed training parameters, reward model implementations, and public dataset manifests on arXiv and Hugging Face Papers.