AI-redeneringmodellen trainen met minder rekenkracht

Researchers at JD.com have developed RLSD, a new training method that enables enterprises to build custom AI reasoning models with significantly fewer computational resources. The technique combines reinforcement learning with self-distillation to overcome limitations of existing approaches, which either require expensive large models or provide insufficient feedback during training. Early experiments show RLSD-trained models outperform those using traditional distillation and reinforcement learning methods, potentially lowering both technical and financial barriers for organizations