Perplexity and NVIDIA Claim JOLT+ Beats GRPO by 9.9 Points on AppWorld
Summary
Perplexity Research and NVIDIA propose JOLT, a self-distillation method that trains one model as both teacher and student. They claim JOLT+ beats GRPO by 4.2 percentage points on MATH and 9.9 points on the AppWorld state-test pass rate. On MATH, JOLT uses 6.5 times fewer processed training tokens at matched accuracy.
Key Points
- Perplexity Research and NVIDIA report JOLT+ reaches 10.96% pass@1 on Terminal-Bench 2.1.
- Perplexity Research and NVIDIA call a teacher reward-consistent when its distillation update matches the student's policy gradient up to positive scaling.
- Perplexity Research says it will open-source the JOLT code soon.