Perplexity and NVIDIA researchers have introduced JOLT, a joint teacher-student reinforcement learning method designed to improve language model post-training efficiency. The technique uses a single shared model that simultaneously acts as both teacher and student, eliminating the need for a separate teacher model during training.
In this setup, the teacher receives privileged information while the student sees only the task, and both roles are trained concurrently. Tested across various benchmarks including GSM8K, MATH, LiveCodeBench v6, AppWorld, and Terminal-Bench 2.1 using models such as Qwen3 and DeepSeek-V2-Lite, the enhanced JOLT+ method achieved the highest mean score in every setting. On the MATH benchmark, JOLT matched GRPO accuracy while using 6.5 times fewer processed training tokens.
This training approach matters because it addresses sparse outcome rewards and compute bottlenecks in reinforcement learning, offering significant efficiency gains for teams running post-training optimization.