Skip to content

Perplexity Introduces JOLT Post-Training Method

Perplexity

JOLT matched GRPO accuracy on the MATH benchmark using 6.5x fewer training tokens.

Key points

  • JOLT uses a single model as both teacher and student, supplying per-token feedback rather than a single reward per attempt.
  • The teacher model receives extra context like verified solutions and is trained on outcome rewards with a penalty for diverging from student predictions.
  • JOLT+ outperformed GRPO across five benchmarks covering math, coding, tool use, and terminal tasks, including a 9.9-point gain on AppWorld.

Why it matters

Model developers can reduce post-training compute requirements while improving accuracy across multi-step reasoning and tool-use tasks.

Sources