Perplexity and NVIDIA researchers have published JOLT, a training method that jointly trains a teacher and a student model to improve reasoning and agentic performance, Perplexity said on 21 October 2026.
The work targets a weakness of reinforcement learning in post-training: a single outcome reward gives a sparse signal that does not show which intermediate steps helped. A common fix, on-policy distillation, uses a stronger teacher to give token-by-token feedback.
The researchers argue that privileged information can make a teacher better at solving a task without making it better at teaching. JOLT therefore trains the teacher relative to the student's current behaviour.
With student outcome rewards, the JOLT+ variant reported absolute gains over GRPO of 1.5 percentage points on GSM8K, 4.2 on MATH, 3.0 on LCBv6, 9.9 on AppWorld and 2.7 on Terminal-Bench 2.1 pass@1. On MATH, JOLT used 13 times fewer completion tokens and 6.5 times fewer processed training tokens than GRPO at matched accuracy.
"We will open-source the code soon," the authors wrote.
The results are the authors' own and had not been independently replicated at publication.