AI

Perplexity Announces JOLT Post-Training Method for More Efficient Reasoning

Wednesday, October 21, 2026Read Original

Details

  • Perplexity introduced JOLT, or Joint On-policy Learning and Teaching, a post-training method that uses one model as both teacher and student.
  • The teacher receives privileged information, such as verified solutions, while the student learns without that information.
  • Unlike standard reinforcement learning methods such as GRPO, which typically provide one reward per attempt, JOLT supplies dense per-token feedback indicating which steps were useful.
  • JOLT jointly trains the teacher using outcome rewards and a penalty for diverging from the student, while training only the student through the teacher’s guidance.
  • On the MATH benchmark, JOLT matched GRPO’s accuracy using 6.5 times fewer processed training tokens at a matched accuracy level.
  • JOLT+ outperformed GRPO across five evaluations spanning mathematics, coding, tool use, and terminal tasks.
  • Reported gains included 1.5 percentage points on GSM8K, 4.2 points on MATH, 3.0 points on LiveCodeBench v6, and 9.9 points on AppWorld.
  • The results suggest that jointly aligned teacher and student training can improve both reasoning quality and post-training efficiency.

Impact

JOLT targets a central cost problem in reasoning-model development: reinforcement learning can require many sampled attempts while offering limited guidance about intermediate steps. Its reported efficiency gains could lower the compute needed to improve models for coding and agentic tasks, while its shared teacher-student design builds on Perplexity’s earlier self-distillation work. The method appears to strengthen Perplexity’s research position in post-training, although practical deployment benefits will depend on whether the gains transfer to larger models and production workloads.

Rift Dispatch