The Pulse
JOLT Reports 6.5× Fewer Training Tokens on MATH
Perplexity researchers introduce JOLT, a method that trains one language model as both teacher and student. On MATH, they report matching GRPO’s accuracy with 6.5 times fewer processed training tokens, using a fitted curve for JOLT’s compar
AI.info Team · Written with AI · fact-checked ·
“Consequently, even a higher-performing teacher can provide guidance that degrades student performance.”
Randy Ardywibowo, Arnav Dalal and Jiantao Jiao, authors of the paper
One model takes both sides of the lesson
Perplexity researchers introduced JOLT, short for Joint On-Policy Learning and Teaching, in a research post published October 21. The method trains one model to act as both a teacher, given privileged information such as a verified solution, and a student, which attempts the task without that information.
The approach targets a weakness in outcome-based reinforcement learning: a model may receive a reward only after completing a long task, leaving it little information about which intermediate decisions helped. On-policy distillation offers denser, token-by-token guidance from a teacher. But, the authors argue, a teacher’s solution can rely on information or shortcuts that the student cannot use.
JOLT trains the two roles jointly. The student learns from the teacher’s token-level guidance on the student’s own attempts. The teacher, meanwhile, receives outcome rewards and a penalty for straying too far from the student’s behavior. Perplexity calls the aim “meet the student where they are”: make the teacher’s feedback useful to the model receiving it, rather than optimizing the teacher’s task performance alone.
The 6.5× comparison uses a fitted curve
The clearest efficiency result comes from the MATH benchmark, where the researchers trained Qwen3-0.6B and gave the teacher worked reference solutions. At a matched accuracy of 63.44%, GRPO’s observed checkpoint used 8.40 million processed training tokens. JOLT’s fitted curve reaches the same accuracy at 1.30 million tokens, which the authors report as 6.5 times fewer.
That comparison is not between two observed checkpoints: the GRPO figure comes from a checkpoint, while JOLT’s comes from a curve fitted to results across training budgets. The processed-token measure includes unpadded prompts, privileged context, optimization passes and teacher-student cross-scoring, not just model-generated answers.
The researchers also compare completion tokens, a narrower measure that counts student completions for GRPO and OPSD and both teacher and student completions for JOLT, but excludes prompts, privileged context and extra optimization passes. On that basis, they report about 13 times fewer completion tokens at the same MATH accuracy. Across a range of target accuracies, the post reports JOLT used 9.5–14.4 times fewer completion tokens and 4.3–7.4 times fewer processed training tokens.
JOLT and JOLT+ have different results
The distinction between JOLT and JOLT+ matters. JOLT trains the teacher and student together, without giving the student direct outcome rewards. In the researchers’ tests, its mean accuracy exceeded GRPO’s on GSM8K, MATH and LiveCodeBench v6, but fell below GRPO and π-Distill on AppWorld.
JOLT+ adds outcome rewards from the student’s own attempts. Perplexity reports that it had the highest mean score among the tested methods on every benchmark setting. Against GRPO, the reported gains were 1.5 percentage points on GSM8K, 4.2 on MATH, 3.0 on LiveCodeBench v6 and 9.9 on AppWorld. On Terminal-Bench 2.1, JOLT+ reached a 10.96% pass rate at pass@1, a 2.7-point gain over GRPO.
The authors flag a gap in the efficiency test
The paper explains that the ideal teacher would guide the student in a direction aligned with the student’s own reward gradient. In practice, the student’s continuation advantages—the expected value of its next actions—are generally unavailable. JOLT therefore uses centered returns from teacher-generated attempts as a substitute. The authors explicitly caution that the substitution is not exact: teacher returns need not match the student’s continuation advantages.
That caveat, along with the fitted-curve comparison, puts the headline efficiency result in context. The reported benchmarks show promising results for the tested models and tasks, while the paper does not establish that the same savings will hold across other model sizes or training settings. Perplexity says it plans to open-source the code; the October 21 post does not announce a release date.