Studying an agent is easier when you can inspect how its tasks and rewards were constructed.
Axion 1.0’s preprint, submitted November 2 and revised November 3, describes an agent-training pipeline built around two supervised fine-tuning stages and reinforcement learning. The authors include researchers from ByteDance and several universities. Their ByteDance repository exposes code for data construction, training, and evaluation.
The stages teach different kinds of work
The first supervised stage focuses on reasoning tasks that can be solved without retrieving outside evidence. The second introduces tasks that need tools. Reinforcement learning then trains longer interactions with search, retrieved pages, and computation.
For tool-dependent tasks, the authors construct questions from public structured data and calculate reference answers. They withhold the construction-time tables during solving. This design rewards finding and combining evidence rather than copying a hidden answer from the training pipeline.
Released code and a checkpoint are different artifacts
The repository includes an Apache-2.0 code license and links training datasets. It also links a 35B community reproduction, explicitly identifying it as a community release. Readers should not infer that this link supplies the official checkpoint evaluated in the paper.
That distinction creates three separate adoption paths. A researcher can inspect the recipe, attempt a training reproduction, or evaluate a community checkpoint. Each path has a different artifact trail. A result from one does not automatically reproduce the paper’s model.
Keep benchmark conditions with the numbers
The authors report results across scientific reasoning, information seeking, and other agent tasks. Their repository distinguishes baselines they evaluated from scores taken from other reports. It also notes restricted website access for its HLE result. These are author-reported findings from a preprint, not an independently reproduced ranking.
A useful reproduction record would preserve the base model, dataset revisions, tool permissions, harness, and stopping rules. Record the exact checkpoint too. A shared training recipe reduces some unknowns, but hardware, data filtering, and agent execution can still change the outcome.
Use the recipe to frame a narrower experiment
I would start by choosing one training-stage claim and its corresponding evaluation rather than trying to validate the entire leaderboard. For example, compare a tool-dependent task before and after the second supervised stage while keeping the tool setup fixed. This is a proposed experiment; we have not trained or evaluated Axion.
Our DeepSeek research-prototype coverage makes a related distinction between an inspectable artifact and a finished hosted service. Which part of an agent recipe is hardest for your team to reproduce: the data, the reward, or the tool environment?
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
