Salesforce study finds 42% compute savings by distilling RL-trained experts

DiarioBitcoin

Factuality

Ownership

A preprint by Salesforce AI Research suggests that training expert models separately and then distilling them into a single system can match the performance of joint reinforcement learning with 42% less computation in a math and code test. The result is promising, although it depends on a specific configuration and does not turn distillation into a universal solution.