Perplexity and NVIDIA researchers introduce JOLT, a method that jointly trains teacher and student rolesMachine translation
Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.
JOLT trains one language model in both teacher and student roles. On MATH, the research team estimated that it needed 6.5 times fewer processed training tokens than GRPO to reach the same accuracy. Without a student reward, however, JOLT scored below GRPO on AppWorld. The code has not yet been released.
This source permits summary display only.
Read at the original source