Perplexity’s JOLT trains one model as both privileged teacher and student

Perplexity published a research paper and blog post on 21 October 2026 introducing Joint On-Policy Learning and Teaching (JOLT), a post-training method for reasoning, coding and agentic tasks. In standard On-Policy Self-Distillation (OPSD), one model acts as its own teacher by reading privileged context such as a verified answer, reference proof or test log, and the unprivileged student learns to imitate it token by token. Perplexity’s authors show that a stronger teacher can actually hurt the student if it relies on shortcuts that the student cannot reproduce without the answer key.

JOLT fixes that mismatch by training both roles jointly inside a single set of weights. The teacher learns from outcome rewards with a Kullback-Leibler (KL) penalty that keeps its token probabilities close to the student’s current policy, while the student learns from the teacher’s dense distillation signal on the student’s own rollouts. A second variant, JOLT+, adds the student’s own centred outcome rewards to the distillation update.

Across five benchmarks (GSM8K on DeepSeek-V2-Lite, MATH on Qwen3-0.6B, LiveCodeBench-V6 on Qwen3-8B, AppWorld on Qwen3-4B-Instruct and Terminal-Bench 2.1 on Qwen3-30B-A3B-Instruct), JOLT+ records the highest mean score in every setting, beating GRPO by 1.5 percentage points on GSM8K, 4.2 points on MATH, 3.0 points on LCBv6, 9.9 points on AppWorld and 4.5 points at pass@32 on Terminal-Bench 2.1.

On MATH, JOLT reaches 63.44% accuracy using 1.30 million processed training tokens compared with 8.40 million for GRPO, or 6.5 times fewer processed tokens and 13 times fewer completion tokens. Perplexity says it plans to open-source the code later.