Perplexity and NVIDIA unveil JOLT training method for language models

Perplexity and NVIDIA introduced JOLT, a post-training method where a single model acts as both teacher and student. On MATH it matched GRPO accuracy with 6.5x fewer processed training tokens, and JOLT+ beat GRPO by 9.9 points on AppWorld.

• JOLT+ scored 72.09% on AppWorld versus 62.20% for GRPO
• On MATH JOLT matched GRPO accuracy with 6.5x fewer tokens
• Tests covered Qwen3 models from 0.6B to 30B-A3B and DeepSeek-V2-Lite
• On Terminal-Bench 2.1 JOLT+ hit 10.96% pass@1 versus 8.25% for GRPO