Athanor preprint reports autonomous agents matching official Qwen model scores

Researchers report in a preprint that four Claude Opus 5.5 agents autonomously post-trained Qwen3-4B-Base to match or surpass its official instruction-tuned version on average across two evaluation suites. The 104-hour Athanor campaign gave each agent an eight-GPU node and no research advice. Scores rose from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on a held-out suite, compared with 66.5 and 47.9 for Qwen3-4B-Instruct-2507. Agents built training data, modified algorithms and combined models using public data inside a sandbox that barred external teacher models. The gains came from recombining known methods, including supervised training, reinforcement learning, weight merging and self-distillation.