Perplexity and NVIDIA researchers cut RL training tokens 6.5x with JOLT
6.5x fewer tokens
Matched GRPO accuracy on MATH with 6.5x fewer training tokens; AppWorld pass rate rose to 72.09%
Researchers at Perplexity and NVIDIA introduced JOLT, a post-training method that trains one language model as both a privileged-information teacher and a student. On the MATH benchmark it matched GRPO's accuracy while processing 6.5 times fewer training tokens. A JOLT+ variant also lifted AppWorld pass rates to 72.09% from 62.20% for GRPO.
Read the full story at MarkTechPost