Update
Salesforce Study: 42% Compute Savings by Distilling RL-Trained Experts
Salesforce's AI Research team has found a new approach to reduce AI training costs by 42% through the use of distillation techniques. By training several specialists separately and then merging their knowledge into a single model, the researchers matched the performance of joint reinforcement learning on a combined math-and-code setup while using 42% fewer total training FLOPs. The study titled 'Unpacking the Compute Efficiency of Strong-to-Weak OPD against GRPO' was released on November 5, 2026.