A ByteDance-led recipe (with a Duke intern as first author, plus UMD and Buffalo) arguing that a strong agent model needs only three post-training stages: two SFT stages and then RL, with no dedicated mid-training, DPO or on-policy distillation. The first SFT stage builds self-contained reasoning from model-aware synthetic problems. Public question-answer pairs are used only as seeds, composed into new self-contained questions, then adversarially hardened against the base model's own successful solution strategies. The second stage teaches tool-dependent reasoning on automatically verifiable tasks built from real-world data sources and deterministic computation. Asynchronous RL then lengthens the horizon over tools and environments.

Applied to Qwen bases, Axion 1.0-122B (from Qwen3.5-122B-A10B) reports 59.0 on full Humanity's Last Exam with restricted web access, 87.0 on FrontierScience Olympiad and 66.7 on FrontierScience Research. That is above Agents-A1 (47.6 HLE) and above GPT-5.6 Sol on the FrontierScience pair as the Axion team ran it, though below Claude Fable 5's reported 63.8 on HLE. Axion 1.0-35B (from Qwen3.6-35B-A3B) scores 62.2 on SEAL-0, and 4B and 9B variants were trained with SFT only. The release is Apache 2.0 code for data generation, filtering, prompts, both SFT stages, asynchronous RL and grading, plus the training data on HuggingFace. The report calls the models fully open, but no weights were on HuggingFace at filing.

post-trainingrlagenticreasoningresearch