What matters in AI.

Subscribe

Agents

4 AI agents train a raw model to the score of its instruct model

The agents work in a sandbox and use no external teacher models.

Researchers make Athanor, an open system in which AI agents do post-training research on GPUs. A campaign of 104 hours gives the result. 4 Claude Opus 5.5 agents increase the score of Qwen3-4B-Base from 15.9 to 67.3 on a development suite of 6 tasks. The instruct model Qwen3-4B-Instruct-2507 gets 66.5. The authors say that this is a first.

Score, 6-task development suite Ranking bars on one scale from 0: Qwen3-4B-Base after agents 67.3, Qwen3-4B-Instruct-2507 66.5, Qwen3-4B-Base before 15.9. Four Claude Opus 5.5 agents, 104 hours. SCORE, 6-TASK DEVELOPMENT SUITE Qwen3-4B-Base after agents Qwen3-4B-Instruct-2507 Qwen3-4B-Base before 67.3 66.5 15.9 4 Claude Opus 5.5 agents · 104 hours
In 104 hours, four Claude Opus 5.5 agents trained Qwen3-4B-Base to 67.3 on a six-task suite, against 66.5 for the instruct model.

Sources

  1. Athanor: An Agent-Native System for Autonomous Post-Training Researcharxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·16 OCT2026

Posted

Tags

Learn the terms in this story

More in Agents

All Agents news