NVIDIA researchers discuss JOLT, a new teacher-student reinforcement learning method, on a large screen with graphs and data.

Editorial illustration for Perplexity, NVIDIA Researchers Introduce JOLT Joint Teacher-Student RL Method

Perplexity, NVIDIA Researchers Introduce JOLT Joint...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

October 22, 2026 • 3 min read

Researchers at Perplexity and NVIDIA built a reinforcement learning method that asks one model to grade its own homework, literally. The method is called JOLT, short for Joint On-policy Learning and Teaching, and it changes how language models get post-trained after their initial run through data.

The setup splits a single model into two roles. One version acts as a teacher with access to privileged information the student never sees. The other version is the student, learning from that feedback. What makes JOLT different from prior self-distillation approaches is that the teacher itself gets trained during the process, so its guidance doesn't go stale as the student improves.

The payoff shows up in compute. On the MATH benchmark, JOLT matched the accuracy of GRPO, a standard outcome-reward RL method, while processing 6.5 times fewer training tokens. For teams running RL post-training at scale, where GPU hours translate directly into budget, that kind of reduction is the whole pitch. The team tested the method across five models, from Qwen3-0.6B up to Qwen3-30B-A3B-Instruct and DeepSeek-V2-Lite, comparing results against GRPO, OPSD, and π-Distill.

On MATH, the research team fit accuracy curves against training-token budgets. At 63.44% accuracy, GRPO’s checkpoint used 8.40M processed training tokens. JOLT’s fitted curve needed 1.30M, a 6.5x reduction.

Perplexity and NVIDIA Researchers Introduce JOLT: a Joint Teacher-Student Reinforcement Learning Method - MarkTechPost

Why this matters

A 6.5x reduction in training tokens to match GRPO on MATH is the kind of number that should make any team running RL post-training stop and check their own pipelines. Compute is the budget line that actually constrains how often labs can iterate on reasoning models, and JOLT's bet, that a single model can hold both the privileged teacher and the blind student without the usual distillation decay, is a genuinely different lever than just scaling up rollouts. We'd want to see this replicated outside MATH before treating it as a general recipe; one benchmark and one paper from Perplexity and NVIDIA is a result, not a consensus.

The detail worth watching is the KL penalty keeping the teacher tethered to the student. That's the mechanism preventing the usual failure mode where the teacher's feedback drifts into something the student can't use. If that holds up on harder, messier tasks than MATH, this becomes a real alternative to GRPO rather than a one-off efficiency trick.

Worth tracking whether other labs reproduce the token savings on different domains.