Perplexity and NVIDIA's JOLT matches GRPO on MATH with 6.5x fewer tokens

Perplexity and NVIDIA researchers have introduced JOLT, a post-training method that trains a single model as both teacher and student at the same time. On the MATH benchmark, the researchers report that JOLT matched the accuracy of GRPO, the reinforcement learning approach widely used for reasoning models, while processing 6.5 times fewer training tokens.

The number that matters here is the token count. Training tokens are a rough proxy for compute and cloud bills, so a method that reaches the same accuracy on a fraction of them can translate into lower costs for teams fine-tuning models for math, code or other reasoning-heavy work. The gain is measured against one benchmark, one baseline and the researchers' own experiments, so it is an early signal rather than a settled rule.

The approach departs from the usual setup, where a larger teacher model guides a separate student. Training both roles jointly means the system learns alongside its own feedback loop, which removes the need to host and run a separate teacher during training. That simplifies the pipeline, though it also raises questions about stability that a single benchmark cannot answer.

For companies building or customizing their own reasoning models, the takeaway is that post-training efficiency is still moving fast. Teams that rely on GRPO-style training should watch whether JOLT holds up on wider tasks and on other organizations' hardware. Its open questions are the ones any promising research result faces: reproducibility, scale and what happens beyond the benchmark it was tested on.