Perplexity and NVIDIA researchers announced JOLT, a joint teacher‑student reinforcement learning method. The approach trains a single model to act as both teacher and student after the main training phase. On the MATH benchmark, JOLT reached the same accuracy as the GRPO method while using 6.5 times fewer training tokens. The result shows a way to improve efficiency without changing the model architecture.
Why it matters
Researchers can fine‑tune models more cheaply, making high‑performance math reasoning more accessible to labs with limited compute budgets.
