Training a large language model to reason well is expensive. Perplexity Research thinks it has found a way to make that process considerably cheaper.
The company, working with NVIDIA, introduced a method called JOLT, short for Joint On-Policy Learning and Teaching, on October 21, 2026. On the MATH benchmark, JOLT reached the same accuracy as Group Relative Policy Optimization (GRPO) while using 6.5 times fewer training tokens.
That matters because GRPO is already considered one of the more efficient options in the field.
One model, two jobs
A single model plays two roles at once: it teaches, and it learns.
In JOLT, the teacher role gets access to what the researchers call privileged information, such as verified solutions. It uses that extra context to give the student role specific, actionable feedback. The student absorbs that guidance as it trains.
The technical name for this is dense on-policy distillation. “On-policy” means the model learns from its own attempts rather than someone else’s, and “dense” means feedback arrives throughout the work instead of only at the end.
The method combines two ingredients: self-distillation, where a model learns from a better-informed version of itself, and reward-consistent teacher training. The second piece keeps the teacher’s guidance aligned with what the training process is actually trying to reward.
The research paper behind the work carries the title “A Good Self-Teacher Meets the Student Where They Are: Joint On-Policy Learning and Teaching.” It was submitted around early October 2026.
The numbers beyond MATH
The 6.5x token savings on MATH is the headline figure. It is not the only one.
An upgraded variant, JOLT+, was tested across a broader set of benchmarks. On AppWorld, it delivered a 9.9-point accuracy gain over GRPO.
The experiments were not limited to one architecture. Researchers ran JOLT on several base models, including DeepSeek-V2-Lite and multiple sizes of Qwen3.
Why sparse rewards have been a problem
Post-training has leaned heavily on reinforcement learning driven by outcome rewards. The model attempts a task, and it gets a signal based on whether the final result was right.
GRPO earned its place in this toolkit as an efficient, critic-free way to estimate how much better or worse an attempt was relative to others.
The weakness is the feedback itself. Outcome rewards are sparse: the model learns it failed, but not where or why. That sparsity has limited how effectively models can learn, especially on long reasoning chains where a single early mistake can sink everything downstream. JOLT’s dense feedback is designed to fill in those missing margin notes.
What this means for AI developers
The most direct implication is cost. Tokens processed during training translate into compute, and compute translates into money and time.
The NVIDIA involvement is worth noting on its own. A chipmaker co-developed the method alongside Perplexity Research.
The AppWorld result may end up being the more consequential number. Long-horizon, multistep tasks sit at the center of the industry’s push toward AI agents, and JOLT+ delivered a 9.9-point gain specifically on this class of benchmark, where reinforcement learning has historically struggled.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.