Perplexity Made A Model Teach Itself By Playing Both Roles

JOLT stands for Joint On-policy Learning and Teaching, and it trains one model in two roles at once. The teacher is given extra information the student does not see — verified solutions, for example — and is trained to produce feedback the student can actually use. The student learns from that feedback on its own outputs. Perplexity reports improvements on math, coding and agent tasks, without publishing specific figures.

The non-obvious idea is in what the teacher is optimised for. Most distillation trains a student to imitate a fixed stronger teacher, which caps the student at roughly the teacher's level. JOLT instead rewards the teacher for producing feedback that improves the student — a different objective, and anyone who has worked with a brilliant colleague who could not explain anything knows the two come apart. Because both roles are the same model trained together, the teacher's advantage is informational rather than architectural: not a bigger model, the same model with privileged access during training.

The missing numbers matter. "Gains in math, coding and agent tasks" with no figures and no stated baseline is not a result you can act on, and training-method papers are where selective reporting is easiest because the authors choose the comparison. The questions that decide whether JOLT is interesting: gains against what — standard RL fine-tuning, or no post-training at all? What is the compute cost of training both roles jointly? And does it hold at frontier scale, or only where the student has obvious headroom? Read it as cost engineering with a paper attached, which is not a criticism — Perplexity runs models at query volumes where inference cost dominates, and that constraint makes the work likely to be real rather than decorative.

Sources

The Neuron daily digest