New AI training method cuts compute costs

Researchers have developed a novel AI training paradigm called Reinforcement Learning with Verifiable Rewards with Self-Distillation (RLSD). This technique significantly lowers the computational and financial barriers for enterprises looking to build custom reasoning models. RLSD combines the reliable outcome tracking of reinforcement learning with the detailed feedback of self-distillation, outperforming traditional methods in experiments. It addresses the limitations of existing approaches like sparse feedback in reinforcement learning and high overhead in distillation, enabling more efficient and cost-effective development of specialized AI.