Perplexity, the AI search company, has described a training method that has a single language model play both teacher and student, and says it beats standard reinforcement learning on math, coding and agent tasks. The method is called JOLT, an acronym for joint learning and teaching on the model’s own attempts. Perplexity published it on its research blog, and every result below is the company’s own.
The problem it targets is familiar to anyone who has trained a model with reinforcement learning. The model attempts a task many times, gets one score per attempt, and drifts toward whatever scored well. That single score cannot say which step in a long attempt helped. On hard agent tasks, where attempts run long and wins are rare, the model gets very little to learn from.
A popular fix is distillation, where a stronger model reviews the student’s attempts word by word. Perplexity argues that being good at a task does not make a model good at explaining it. A teacher’s advice can sit far beyond what the student can do yet, so the student picks up little from it.
JOLT builds on a variant in which one model takes both seats. As the teacher, it sees extra information the student never gets, such as a verified solution. As the student, it works without that help. The older variant only trained the student to copy the teacher. JOLT also trains the teacher, and rewards it when its guidance pushes the student the same way the student’s own reward would. The teacher is graded on whether its lessons work for this particular student. A second version, JOLT+, also hands the student its own outcome scores.
Perplexity tested five setups, each with a different open base model from DeepSeek or Qwen, ranging from 0.6 billion to 30 billion parameters. They cover grade-school math (GSM8K), a harder math set (MATH), 131 coding problems from LiveCodeBench, AppWorld, a test of agents that operate apps, and Terminal-Bench 2.1, which tests agents working in a command line. The comparison methods were GRPO, a standard reinforcement learning recipe, and two distillation approaches.
Perplexity calls GRPO the strongest baseline. Against it, JOLT+ gained 1.5 percentage points on GSM8K, 4.2 on MATH, 3.0 on the coding set and 9.9 on AppWorld. On Terminal-Bench 2.1 it gained 2.7 points on the first attempt, and its score rose from 10.96 percent on one try to 34.83 percent when allowed 32.
Plain JOLT, without the student’s own reward, did less well. It topped GRPO on GSM8K, MATH and the coding set. It fell below both GRPO and the rival distillation method on AppWorld, and stayed under GRPO on Terminal-Bench. The post is open about this, and it means the strongest numbers belong to the version that still leans on ordinary rewards.
The efficiency claim is the one with the most commercial weight. On MATH, to reach 63.44 percent accuracy, GRPO’s measured checkpoint used 8.40 million training tokens. A curve fitted to JOLT’s runs puts it at 1.30 million, which is 6.5 times fewer. Counting only completion tokens, the gap grows to about 13 times. JOLT also levelled off near 72 percent accuracy, against about 70 for GRPO and 62 for the older self-distillation method.
Perplexity says the code will be released as open source soon, so outsiders cannot yet rerun any of this. Training compute is a large line in any post-training budget, and a method that reaches the same accuracy on a fraction of the tokens would change that line. The 6.5 times figure rests on one math benchmark and the smallest model in the study, though, so it is a reason to test once the code ships and not a number to plan around.
Reported by Perplexity on its research blog, in the post “A good teacher meets the student where they are”; the page carries no publication date.
