Axion 1.0: An Open Agent Model Built with Three Training Stages

Introduction

An agent model must do more than produce a correct answer. It needs to decompose goals, retrieve information, call tools, execute multi-step plans, and revise its behavior when the environment provides new feedback. Axion 1.0, introduced by Yuqi Jia and collaborators, explores whether these capabilities can be developed with a relatively compact post-training recipe rather than a long sequence of specialized stages.

A three-stage recipe

The proposed process combines two lightweight supervised fine-tuning stages with reinforcement learning:

• Build self-contained reasoning first. The first stage targets scientific problems that can be solved without external tools. The authors construct challenging examples with awareness of the base model’s own solution behavior, then adversarially strengthen successful problems to make the reasoning demand more robust.
• Add tool-dependent investigation. The second stage introduces tasks that require retrieving information from real-world sources, performing deterministic computation, or using external tools. The pipeline is designed to generate diverse tasks with automatic verification, reducing dependence on expensive subjective labeling.
• Improve interaction through reinforcement learning. The final stage trains on agent trajectories, with the goal of strengthening long-horizon reasoning and repeated interaction with tools and environments.

The project explicitly says that it does not use a dedicated mid-training phase, direct preference optimization, or on-policy distillation. Its public description also cites approximately 18,000 SFT examples and mentions 35B and 122B versions of Axion 1.0.

Results and what is released

The reported evaluation suite covers HLE, BioMysteryBench, SEAL-0, HiPhO, and the Olympiad and Research splits of FrontierScience. For the 35B reproduction checkpoint, the listed scores are 52.8, 35.3, 60.4, 45.9, 85.0, and 55.0 respectively. These figures should be interpreted in the context of the project’s evaluation setup, including prompts and tool environments; independent reproduction would be needed to establish comparability across implementations.

Axion 1.0 is presented as open beyond its weights. The release includes data construction and generation procedures, filtering and verification, prompts, training code, training data, and a 35B reproduction model. That checkpoint is released under Apache 2.0, giving researchers a practical starting point for testing and extending the method.

Why it matters

The project’s main contribution is a clear capability progression: solve problems independently, investigate with tools, and then learn through interaction. If the released pipeline proves reproducible, it could help researchers isolate the roles of task construction, automatic verifiers, and reinforcement learning in agent training.

There are also important caveats. A simple recipe does not necessarily mean low engineering cost. The quality of generated tasks, the coverage of verifiers, the reliability of tool environments, and the stability of long-horizon reinforcement learning can all affect outcomes. Axion 1.0 is therefore best viewed as an open research baseline and a reproducible direction, rather than a final measure of general-purpose agent intelligence.

Source: Hugging Face Daily Papers