RSI Arena: AI Agents Self-Train Models with Mixed Results

Researchers from Bake AI, Notre Dame, UCLA and NVIDIA released RSI Arena, where 10 frontier AI agents autonomously fine-tune models. Eight improved with human feedback and the best closed 76% of the baseline gap, but gains diverged sharply: DeepSeek's model gained 172 human-eval points with only +4 on benchmarks, while MiniMax's gained 68 human points but dropped 18 on benchmarks.

2026-10-28 ~ 2026-10-29 · 3 related posts

• RSI Arena: 10 AI Agents Autonomously Post-Train Models, Best Closes 76% of Gap — RSIArena · 2026-10-28
• RSI Arena paper: 10 AI agents autonomously post-train models, 8 improve with human feedback — teortaxesTex · 2026-10-29
• SFT benchmaxxing backfires: DeepSeek-tuned model gains 172 human rating points — teortaxesTex · 2026-10-29