Athanor: An Agent-Native System for Autonomous Post-Training Research
Athanor agents autonomously post-trained Qwen3-4B-Base to match official Qwen3-4B-Instruct on held-out tasks.
Athanor is an open agent-native system for autonomous post-training research on real GPUs, with frozen evaluation, managed jobs, decontaminated data, and a hidden held-out suite. In a 104-hour run, four Claude Opus 5.5 agents, each on an eight-GPU node and barred from external teacher models, improved Qwen3-4B-Base from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on held-out tasks. Those scores match or slightly exceed the official Qwen3-4B-Instruct-2507 results of 66.5 and 47.9. The agents reached that level by combining SFT on public reasoning data, an agent-introduced RL rollout change, cross-agent weight merges, and self-distillation.
- Four Claude Opus 5.5 agents ran a 104-hour post-training campaign with no research advice.
- Qwen3-4B-Base rose from 15.9 to 67.3 on six development tasks, versus 66.5 for official instruct.
- Held-out performance rose from 9.2 to 48.2, compared with 47.9 for Qwen3-4B-Instruct-2507.
- Agents recombined public-data SFT, RL, weight merges, and self-distillation without external teachers.
- Products
- Athanor
From the original article
1 min read at huggingface.co
Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence. We study a weaker form that can be measured today: frontier agents post-training a small raw base model with public data, as a test of their research potential. We build Athanor, an open agent-native system for autonomous post-training research (AutoPostTrain) on real GPUs.
By Yufa Zhou, Yichen Wang, Jiakun Fan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.