On-Policy Delta Distillation
Teaching AI to learn by copying what the teacher learned, not what it says
Researchers developed a new way to train reasoning AI models by having them copy the specific improvements a teacher model gained during training, rather than copying the teacher's final outputs. This approach, called On-Policy Delta Distillation, consistently outperformed standard methods across math, science, and coding tasks while requiring significantly less training time.
Most AI training is computationally expensive and relies on human feedback at scale. This method cuts training time substantially while maintaining strong performance, making it cheaper and faster to build AI systems that can reason through complex problems. The approach works because it focuses on transferring the actual reasoning improvements rather than surface-level mimicry, making the training signal more direct and effective.