Google just fused two opposite training methods into one system — Supervised Reinforcement Learning (SRL) — and made small 7B models earn intelligence with dense, step-wise rewards. In tests on s1K 1.1 and code tasks, SRL beat standard SFT and held up when no rollout was correct. The winning recipe: SRL first, then RLVR. Meanwhile, Gemini 2.0’s multi-agent “AI co-scientist” helped researchers surface drug candidates for liver fibrosis and independently reconstructed a decade-long discovery about phage tail piracy in bacteria—in days.
Credit to : AI Revolution
Please support our Sponsors here :