Gemini 4 Leak Unconfirmed as Google DeepMind Publishes Dream-RSI Paper

Amid an unverified Gemini 4 benchmark leak, Google DeepMind published the Dream-RSI paper (arXiv 2609.14858) on September 14 about recursive agent self-improvement. The method cuts model calls 162x on the Lasso path solver task (51,200 vs 317) without fine-tuning weights.

• Dream-RSI cut calls from 51,200 to 317 on Lasso path solver
• KernelBench VGG16 needed 2.43x fewer generations at same quality
• Method does not change model weights; used Gemini-3.1-Pro and Gemini-3.7-Flash
• Leaked Gemini 4 benchmarks have no confirmed source