arXivAnalysisResearchers demonstrate autonomous model training with Athanor agents
What happened.Researchers released Athanor, an open platform where autonomous software agents trained a raw 4-billion-parameter base model to match official instruction checkpoints.
Why it matters.Post-training traditionally requires months of manual trial-and-error by research teams. Autonomous agent clusters can compress fine-tuning iteration cycles from quarters to days.
The technical specifications:
- Autonomous campaign: Four Claude Opus 5.5 agents ran for 104 hours on eight-GPU clusters without human advice to optimize fine-tuning workflows.
- Benchmark gains: Boosted evaluation accuracy from 15.9 to 67.3 on core tasks and 9.2 to 48.2 on held-out benchmarks, matching official releases.
- Method recombination: Agents autonomously discovered rollout adjustments, executed inter-agent weight merges, and ran multi-round self-distillation.