RSI Arena: 10 AI Agents Autonomously Post-Train Models, Best Closes 76% of Gap

RSIArena · hf · 2026-10-28

Bake AI, Notre Dame, UCLA, NVIDIA and others present RSI Arena, a two-stage study of autonomous post-training toward recursive self-improvement.

Setup: 10 frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lighting 30B-A3B with 1,000 GPU-hours and $300 API credit on a shared 64-GPU cluster, autonomously picking data, writing training procedures and choosing checkpoints.
Stage 1: all ten agents produce working assistants; the strongest (trained by GPT-6 Astra) closes a median 76% of the base-to-Instruct gap across 12 benchmarks, with human-preference ratings near NVIDIA Instruct after statistical adjustment.
Stage 2: agents receive comparison records and votes on their Stage 1 model plus three more training days; most achieve preference gains despite modest benchmark improvements.

More from AGI Musings