Many agent-model recipes rely on a stack of training stages that makes it hard to tell which parts actually matter. Axion 1.0 tests a leaner alternative: two targeted supervised fine-tuning stages, then reinforcement learning. Its notable contribution is not just fewer stages, but a way to build tool-use tasks with answers verified by deterministic computation—making the training signal more reproducible and the full recipe easier to inspect.
Key Findings
- The first SFT stage strengthens self-contained scientific and mathematical reasoning by making problems harder based on what the base model can already solve. This targets capability gaps rather than simply scaling a static question set.
- The second stage uses tasks grounded in real-world databases, with programs selecting entities and computing reference answers before questions are written. Solvers must retrieve evidence with tools, so success depends on both reasoning and interaction.
- Asynchronous GRPO then trains longer tool-mediated interactions, with partial credit for verified intermediate progress. The authors report Axion 1.0-122B scores of 59.0 on HLE under restricted website access, 87.0 on FrontierScience Olympiad, and 66.7 on FrontierScience Research.
- The release covers data construction, filtering, prompts, training, and evaluation—not only model weights. That makes the work especially useful for studying and adapting the development process, although the weights are described as forthcoming.
Method and Trade-offs
The reported recipe omits dedicated mid-training, direct preference optimization, and on-policy distillation. The paper also reports SFT-only 4B and 9B variants, while its headline SFT-plus-RL models use Qwen-based 35B and 122B MoE backbones. Benchmark results are informative but should be read with evaluation conditions in mind: some baseline scores were measured by the authors, and the HLE result specifies restricted website access.
Who It’s For
A good fit for researchers who want to reproduce agent training, study verifiable tool-use data, or compare a compact SFT-plus-RL pipeline with more elaborate recipes. Look elsewhere if you need immediately downloadable model weights or evidence that the method transfers unchanged beyond the reported backbones and benchmarks.