Dan Austin has open-sourced a system where an RL-trained Qwen3.6-35B-A3B agent autonomously writes and submits complete training jobs for smaller Qwen models (0.6B or 1.7B). The agent utilizes the prime-rl harness to define environments, rewards, datasets, and hyperparameters, which are then executed on Runpod GPUs.
- The outer RL loop uses Tinker with LoRA and GRPO, rewarding the agent based on the inner model's improvement on a hidden evaluation.
- Over 54 outer-loop steps involving approximately 1,750 GPU training jobs, the episode reward climbed from ~0.0 to a peak of ~0.63.
- The agent learned to prefer the 1.7B base model (usage rose from 42% to 95%) and effectively utilized the hyperparameter configuration space.
- Skill transfer was observed in a held-out task family, with mean reward increasing from 0.399 to 0.545 before plateauing.
- The entire experiment cost approximately $1.3k, with each inner training job costing between $0.13 and $0.30.
The project demonstrates that AI systems capable of improving other AI systems are becoming more accessible, requiring significant process debugging rather than just policy tuning.