Axion 1.0 Trains Tool-Using Agents with Two SFT Stages and GRPO
This is a Plain English Papers summary of a research paper called Axion 1.0 Trains Tool-Using Agents with Two SFT Stages and GRPO. If you like these kinds of analyses, you can find more AI and machine-learning research o
This is a Plain English Papers summary of a research paper called Axion 1.0 Trains Tool-Using Agents with Two SFT Stages and GRPO. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
Axion 1.0’s three-stage recipe for agent training
Axion 1.0 trains agent foundation models with two supervised fine-tuning stages followed by reinforcement learning, omitting dedicated mid-training, direct preference optimization, and on-policy distillation. The central claim is that a compact pipeline can still support reasoning, tool use, and long-horizon interaction when its training tasks are carefully constructed and verified.
The first SFT stage targets self-contained reasoning: seed questions from subjects including mathematics, physics, and biology are transformed into harder problems, then strengthened when the base model solves them. The authors use composition, parameterization, range shifts, and case analysis, and independently verify each rewritten answer before retaining it.
The second SFT stage teaches tool-dependent reasoning using tasks built from public structured data. Programs filter records, select a unique entity, and compute a reference answer; an LLM verbalizes the task, while checks reject ambiguous selections and questions that can be answered correctly without retrieval. The construction database is withheld during solving, so the model must search, inspect pages, and calculate.
From verified demonstrations to long-horizon tool use...
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.