# Perplexity's JOLT Study Estimates 6.5-Fold Lower Training Token Use

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/37631
Published: 2026-10-22T08:05:12.000Z
Updated: 2026-10-22T08:05:12.000Z
Section: Technology

> At 63.44% accuracy on MATH, JOLT's fitted budget was 1.30 million processed tokens, compared with 8.40 million at an observed GRPO checkpoint.

Perplexity and NVIDIA researchers estimate JOLT would need 6.5 times fewer processed training tokens than Group Relative Policy Optimization (GRPO) to reach the same math benchmark score, indicating a possible efficiency gain.

The comparison used Qwen3-0.6B and targeted 63.44% accuracy on MATH. GRPO reached that score at an observed checkpoint after 8.40 million processed tokens. JOLT's corresponding budget of 1.30 million tokens was inferred from a fitted accuracy curve.

That makes the 6.5-fold difference an estimate at one accuracy target. It does not show that a JOLT training run takes 6.5 times less time or costs 6.5 times less money. The researchers did not calculate a confidence interval for the token-budget ratio.

JOLT trains one model in teacher and student roles. In the math experiment, the teacher had access to worked reference solutions, while the student attempted problems without them. The teacher scored tokens in the student's completed responses, providing a learning signal tied to individual parts of each answer rather than only a final result.

The processed-token measure covers more than generated answers. It includes prompts and completions used during training and scoring, along with context available only to the teacher and repeated optimization passes. Some scoring passes count even though they do not update the model's weights.

Counting completion tokens alone, JOLT's fitted budget was 617,000, compared with 8.04 million at GRPO's observed checkpoint, a roughly 13-fold difference. JOLT's figure includes completions from both teacher and student; GRPO's counts student completions.

The paper was submitted Oct. 7, and Perplexity presented the method Oct. 21. The approach addresses a limitation of training that rewards or penalizes an entire response without identifying which parts helped the model arrive at its answer.

A separate variant, JOLT+, adds outcome rewards to student training. On the 57-task AppWorld development test, using Qwen3-4B-Instruct, it achieved a mean state-test pass rate of 72.09%. GRPO scored 62.20%, while base JOLT scored 51.92%.

JOLT+'s 9.89-percentage-point lead over GRPO on AppWorld applies to that variant, not to base JOLT. The AppWorld test and the MATH token comparison measure different results.