Rufus-Air: An Open LLM Post-Training Recipe

What it claims

Rufus-Air post-trains GLM-4.5-Air-Base (106B total / 12B active MoE) through eight serial stages and beats the vendor's own GLM-4.5-Air post-trained release on every reported benchmark except Arena-Hard v2 Creative Writing, while staying competitive with INTELLECT-3 (same base) and Nemotron-3-Super. The report's stated contribution is the recipe's reproducibility: it uses only open-source components (Slime for training, SGLang for serving), fits on 8–32 nodes, and does zero response regeneration — all prompts/responses come from 17 named public datasets used as released, with no in-house distillation teacher or new human annotation.

How it works

The pipeline order is set by exposure to reward hacking, not reward format. Hard verifiable-reward stages (Reasoning RL, Coding RL) run first; softer judge/model-based stages run last (RLHF, the preference judge, is where hacking risk is real). The telling exception: IF RL uses a rubric LLM judge yet runs early, because instruction-following is close to what the model already does, so the judge has little room to be gamed. Difficulty filtering drops prompts the policy already solves (and usually those it never solves), acting as an automatic curriculum. Every later stage updates the same parameters, so gains are not protected — each stage is measured against the checkpoint it starts from, not the base.

SFT is treated as a capability stage, not a warm-up: 9.01M samples / 27.0B assistant-loss tokens across six categories. Sample share and token share diverge sharply — General Agent is 39.6% of samples but 14.2% of tokens, while Math + Coding Agent are 18.8% of samples but 49.1% of tokens (long reasoning traces and multi-turn trajectories).

Evidence

All pass@1 (Arena-Hard rows are win rates), Rufus-Air vs GLM-4.5-Air baseline, same harness:

• IFBench (prompt strict): 76.9 vs 33.6; IFEval 95.4 vs 83.0; Multi-Challenge 65.8 vs 36.0
• Tau2-Telecom: 93.0 vs 32.7; MCP-Atlas 45.3 vs 35.9
• HLE-Verified Gold: 51.1 vs 20.2; Seal-0 51.4 vs 33.3; BrowseComp 37.1 vs 22.7
• LiveCodeBench v6 76.4 vs 59.6; SWE-bench Verified 65.6 vs 50.6
• AIME 25 88.3 vs 84.2; GPQA 75.6 vs 73.9 — small gains where a single stage (Reasoning RL) does the work

The stagewise table (Table 2) attributes each delta to its stage: e.g. IF RL alone moves IFBench +14.0 and Multi-Challenge +24.7; RLHF moves Arena-Hard CW +14.4. SFT actually drops GPQA (68.2, −5.7 vs base) before Reasoning RL recovers it (+5.3).

Limits

• Some conclusions come from training experience, not full ablations, and the authors mark them as such (§6) — this is "between a research paper and an engineering report."
• Loses on Arena-Hard v2 Creative Writing (53.0 vs GLM's 60.3) and Tau2-Airline/Retail vs Nemotron-3-Super.
• Reward-hacking ordering does not protect earlier gains, since later stages overwrite the same weights.
• The infrastructure claims (token-in/token-out rollouts, consistent chat-template handling, sandbox reliability, Rollout Routing Replay for stable RL) are presented as part of the recipe but the fetched text stops before the detailed infra section (§4).
• No licence for the recipe/checkpoints is stated in the fetched text.