Axion 1.0 cuts agent training to three stages: two SFT rounds plus RL, with data pipelines and code released alongside the weights
Synopsis
ByteDance and collaborators released Axion 1.0, a fully open agent foundation model trained with only two lightweight supervised fine-tuning stages followed by reinforcement learning, without dedicated mid-training, direct preference optimization or on-policy distillation. With that compact recipe it stays competitive with leading open models and several frontier commercial models on reasoning, tool use, information seeking and long-horizon interaction benchmarks. The release goes beyond weights to include data construction, generation, filtering, verification, prompts and training code.
Interpretation
Axion 1.0 reduces agent foundation model training to three stages: two supervised fine-tuning rounds followed by one reinforcement learning round, with no dedicated mid-training, direct preference optimization or on-policy distillation. Earlier agent models often stacked continued or mid-training, multiple supervised stages, capability-specific data mixtures, preference optimization and extra post-training steps, and the accumulation made it hard to tell which ingredients were necessary while raising reproduction cost. Across six benchmarks covering reasoning, tool use, information seeking, coding and long-horizon interaction, Axion 1.0 is competitive with leading open models and several frontier commercial models; its 35B and 122B configurations are initialized from Qwen3.6-35B-A3B and Qwen3.5-122B-A10B respectively.
The capability comes mainly from data construction rather than stage stacking: the first supervised round trains self-contained reasoning, the second trains reasoning that depends on tools for external information, and reinforcement learning extends long-horizon interaction. Self-contained reasoning data starts from questions in public course materials, competitions and reasoning datasets; a generator rewrites each into a new question requiring an intermediate result of the seed, the base model attempts it, and problems it solves are strengthened by composition, parameterization, range shift or case analysis. Tool-dependent data is built from structured public databases through filtering, unique-entity selection, deterministic computation and verbalization, with construction-time databases withheld from the solver and closed-book-solvable questions removed; only trajectories with matching reference answers and valid tool calls are kept.
The reinforcement learning stage combines outcome and process rewards, giving partial credit for verified intermediate progress and penalizing trajectories that never submit a final answer. Supervised fine-tuning learns from successful demonstrations but does not directly optimize the model's own decisions when it generates and acts, so group-relative policy optimization follows it. Intermediate targets are checked in dependency order, so a later step earns credit only when all prerequisites pass; RL questions come from the same programmatic pool, four trajectories are sampled per question, and a question is kept only if at least one is correct and the four rewards have nonzero variance.
Perspective
Results cover HLE, the human-difficult subset of BioMysteryBench, SEAL-0, HiPhO and the Olympiad and Research splits of FrontierScience, with models evaluated in a ReAct environment supporting web search, webpage visits, scholarly retrieval and Python execution under benchmark-specific tool restrictions; training-stage and model-size comparisons use a separate 130-question text-only HLE subset whose scores are not directly comparable with the main HLE results.
Externally reported baselines may differ in tool access or execution settings, so cross-source comparisons are indicative rather than strictly controlled; the per-stage checkpoints are sequential comparisons, not component-removal ablations, and RL gains on scientific reasoning benchmarks vary with model size, with the 35B model dipping slightly on FrontierScience Research and HiPhO after RL; routing-replay diagnostics also reflect asynchronous policy staleness and off-policy effects of context condensation, so benchmark differences cannot be attributed to that technique alone.
