RSI Arena enables autonomous model post-training via agents

RSI Arena demonstrates autonomous post-training where agents select data and training procedures using 1,000 GPU-hours and $300 API credit. One agent, driven by GPT-6 Astra, recovered 76% of the base-to-instruct gap on 12 benchmarks, achieving human-preference ratings near NVIDIA Instruct.