Perplexity and NVIDIA Introduce JOLT: Joint Teacher-Student Reinforcement Learning Method

Researchers from Perplexity and NVIDIA have introduced JOLT (Joint On-Policy Learning and Teaching), a new reinforcement learning method that trains student models using a joint on-policy process with a teacher model rather than purely offline distillation. The key innovation is that the teacher and student interact on-policy during training, allowing the student to learn from corrections in contexts it actually generates—rather than only from static teacher demonstrations. This addresses a core weakness of standard knowledge distillation, where student models often fail on distributions that differ from the teacher's demonstrations. For ML engineers working on efficient model training or distillation pipelines, JOLT offers a potentially more data-efficient and capability-preserving path to smaller models. The collaboration between a frontier search AI company and NVIDIA's research team signals growing investment in post-training efficiency research.