Axion 1.0, an open agent foundation model (axion1-0.github.io)

Axion 1.0 has been introduced as an innovative open agent foundation model designed to enhance scientific reasoning capabilities in AI. The model follows a straightforward training approach comprising three key stages: two rounds of supervised fine-tuning paired with reinforcement learning. This method allows Axion to build robust reasoning skills from challenging, model-aware scientific problems, gradually incorporating external information and refining performance through interactive feedback. Notably, it does not rely on complex training processes like direct preference optimization or on-policy distillation, showcasing that powerful agentic capabilities can emerge from a simpler framework. For the AI/ML community, Axion 1.0's significance lies in its demonstration of a viable path to creating capable agents with strong reasoning skills without the need for intricate setups. Evaluated against leading models across several benchmarks, Axion performs remarkably well in areas like scientific reasoning and biological data analysis, with scores exceeding competitors in multiple categories. The model's ability to leverage structured data sources for question construction and effective tool use positions it as a practical resource for future research in agent foundation models, opening avenues for further exploration in AI interaction and problem-solving methodologies.