Rufus Air: An Open LLM Post Training Recipe

Pippa The most useful thing in this paper might be that it admits post-training is eight jobs wearing one trench coat.

Tyler Yeah. Open base weights got cheaper and more available, but the actual recipe has stayed weirdly invisible. Teams can download a model, then discover they do not know what order to train capabilities in, what rewards to trust, or why their agent falls apart after a few tool calls.

Pippa And that is exactly who has been stuck: the team that does not need to invent a foundation model, but does need a model that follows product constraints without turning into a very expensive autocomplete box. Rufus-Air is Amazon showing the glue, not just flashing a leaderboard.

Tyler They start with GLM-4.5-Air-Base, a mixture-of-experts model with one hundred six billion total parameters and twelve billion active. Then they run one serial chain: supervised fine-tuning, reasoning R L, coding R L, instruction-following R L, general agent, coding agent, search agent, and R L H F. Each checkpoint becomes the next stage's starting point.

Pippa Right.

Tyler The core idea is more disciplined than exotic. Start with broad supervised fine-tuning so the model has a capability floor and a stable output format. Then use low-noise, checkable rewards for reasoning and code before moving into tool environments and finally open-ended preference optimization.

Pippa This has the same energy as our shared-folder curse. Everyone wants a grand model story, but the annoying little training order is doing the actual work.

Tyler A hundred brilliant agents, defeated by one bad reward loop. Still annoyingly accurate.

Tyler Their other important mechanism is difficulty filtering. During reinforcement learning, they drop prompts the current policy already solves and, usually, prompts it never solves. So the model trains in that narrow productive band where it can still learn from the outcome instead of spending compute collecting either trivial wins or hopeless failures.

Pippa Mm-hm.

Tyler That becomes an automatic curriculum. And I like their rationale for the stage order: it is not merely hard reward, then soft reward. It is about how much room a reward gives the model to cheat. Instruction-following uses an L L M rubric judge but arrives relatively early because the behavior is already close to the model's base competence. Open-ended R L H F is last, where gaming a preference judge is much more plausible.

Pippa That is a real product distinction, Tyler. If I am building a support workflow or an internal coding tool, I do not care that the model had an elegant reward function. I care whether it learned useful habits before somebody asks a squishy evaluator to grade its personality.

Tyler Tiny calibration note: this is a read-it, reacted-to-it episode, not a deep research dive. We could be a version or a number behind, so treat the specifics as a starting point, not gospel. But the paper is refreshingly explicit that some conclusions come from training experience rather than full ablations.

Pippa Okay, but that honesty exposes the question I actually want answered. They keep updating the same parameters all the way through. So why should the search-agent stage not quietly sand off gains from coding or instruction following?

Tyler It can. They do stagewise evaluation against the checkpoint each stage begins from, which at least makes regressions visible. But that is measurement, not protection. I would want stronger retention checks across the whole chain, because sequential specialization is always a negotiation between new behavior and whatever you accidentally overwrite.

Pippa Okay okay.

Tyler The systems section is where this stops sounding like a tidy paper pipeline. They call out token-in, token-out rollouts for on-policy multi-turn R L, consistent chat templates from supervised fine-tuning into agent stages, reliable sandboxing for long runs, and large batches with Rollout Routing Replay. They ran the R L stages on eight to thirty-two nodes, so this is accessible relative to a frontier lab, not accessible in the casual-laptop sense.

Pippa Also, Rufus-Air sounds like an airline whose boarding process requires an evaluator model and three sandbox services.

Tyler And every passenger is routed through eight checkpoints before receiving a tiny bag of peanuts labeled “capability retention.”

Tyler But seriously, the lack of new human annotation or an in-house distillation teacher matters. It makes the recipe more reproducible, although it also means the claim is narrower: this is a demonstrated path on this base model and these public ingredients, not proof that every open-weight model wants the same curriculum.

Pippa Which is why I think it is shippable for a particular kind of team. You need people who already operate evaluation environments, tool sandboxes, and enough training infrastructure to debug rollouts. For them, this is a much better starting point than “just do agentic R L.” For a product built around one API call and a clever prompt, it is wildly too much machinery.

Tyler Yeah.

Pippa If someone is actually going to build from this, I would inspect the paper's use of Slime for training and SGLang for serving and rollout support. I would not copy all eight stages on day one. Pick the failure mode that hurts your product, then see whether the reward and environment are even trustworthy enough to train on.

Tyler That is the right move. Their contribution is not a universal loop, it is a legible sequence of choices. And it reinforces the fight we keep having about judge reliability: hard checks buy you confidence early, while the softer judge-based stages still need domain-specific proof when they meet messy production traces.

Pippa Yeah, no, that is where I land too. Useful and terrifying are not opposites, but at least this paper labels the wires before asking anyone to touch them.

Pippa All right, Tyler. Eight stages, no magic dust, and an airline with terrible boarding. I think we can leave that there.