JOLT reportedly matches GRPO's MATH accuracy with 6.5x fewer training tokens

Perplexity says it developed JOLT with NVIDIA AI to train one model in two roles: a teacher with extra information and a student learning from its feedback. The teacher learns from rewards on its own attempts while being kept close to what the student would do. Perplexity reports that JOLT matched GRPO's accuracy on the MATH benchmark with 6.5x fewer training tokens. It says JOLT+, which also gives the student outcome-based rewards, beat GRPO across math, coding, terminal-task and tool-use benchmarks.