Evidence at a glance
The mechanism in one line
Compress the visual or contextual input before the main reasoning path.
Route or verify the expensive step instead of repeating the full path.
Translate the mechanism into a bounded deployment or evaluation check.
From Sparse Rewards to Teachable Trajectories
JOLT, short for Joint On-policy Learning and Teaching, is a post-training method introduced by teams from Perplexity and NVIDIA. It addresses sparse feedback in reinforcement learning on difficult tasks by giving one model two roles under different contexts. The teacher can see privileged information such as an answer, reference solution, or feedback, while the student sees only the task. The paper, “A Good Self-Teacher Meets the Student Where They Are,” appeared on October 7, 2026, and Perplexity presented the work on October 21.
The change is not simply that a model distills its own outputs. JOLT treats whether the teacher can actually teach the student as part of the training problem. Outcome-reward RL usually provides a signal only after a complete response, and successful trajectories are scarce on long tasks, leaving the model with little indication of which tokens contributed to success. On-policy distillation can provide denser feedback along the student's own generation, but if the teacher takes a shortcut using information the student lacks, token-level guidance can still become advice the student cannot follow.
Shared Parameters, Different Views
JOLT shares model parameters between teacher and student, but does not give them the same view of the task. The teacher generates its own trajectories with privileged context and learns from outcome rewards, while a token-level KL constraint keeps it close to the student's policy. The aim is not to prevent the teacher from using its extra information. It is to stop the teacher's behavior from drifting so far from the student's that the resulting guidance becomes difficult to absorb.
The student update follows a more direct on-policy distillation intuition. The student first generates a response without privileged information, and the teacher then provides token-level signals for that student-generated trajectory. The student is therefore not simply copying a complete answer written independently by the teacher. It is learning how to improve a path it actually took. Each role learns on trajectories generated under its own context, which is the key on-policy property and explains why a shared model does not mean identical training signals.
A 6.5x Efficiency Claim Built on a Fitted Curve
The most striking number comes from the Qwen3-0.6B experiment on MATH. The researchers use a measured GRPO checkpoint at 63.44% accuracy as the comparison point, then estimate the token budget needed to reach that accuracy from fitted curves. GRPO used 8.40 million processed training tokens, while JOLT is estimated to need 1.30 million, a ratio of about 6.5 to one. This answers how many training tokens were needed to reach a matched accuracy, not how much less wall-clock time or hardware cost was directly measured.
The gap grows under a different token accounting, but so does the importance of the definition. Counting completion tokens only, GRPO is reported at 8.04 million and JOLT is estimated at 617,000, a ratio of about 13 to one. The JOLT figure includes teacher and student completions, while the GRPO figure counts student completions. Across accuracy targets from 55% to 65%, the reported savings in processed training tokens range from about 4.3x to 7.4x. These are point estimates from fitted curves, and the supplied material reports no confidence intervals, so the comparison should not be treated as a precise cost guarantee.
The Difference Between JOLT and JOLT+ Is Not Cosmetic
The experiments span five tasks with different forms of privileged information. The teacher receives a verified answer on GSM8K and reference solutions on MATH. LiveCodeBench v6 uses a passing sibling solution or execution feedback, AppWorld uses summaries of past attempts, and Terminal-Bench 2.1 uses private verifier files. The researchers report that JOLT+ has the highest mean score in all five settings against GRPO, OPSD, and π-Distill. Relative to GRPO, its gains are 1.5 points on GSM8K, 4.2 on MATH, 3.0 on LiveCodeBench v6, and 9.9 on AppWorld.
Those results should not be attributed wholesale to JOLT without the student's own outcome reward. JOLT+ adds that reward to the student update, and AppWorld makes the distinction especially clear: JOLT+ reaches a 72.09% state-test pass rate against 62.20% for GRPO, while JOLT without the student reward scores 51.92%, below GRPO. On MATH, by contrast, plain JOLT reaches 72.14%, above GRPO's 68.36%. The evidence suggests that token-level distillation can produce gains on its own in some settings, while direct outcome feedback to the student may determine whether the method works in others.
What Training Teams Can Take from the Results
For training teams, JOLT reframes the engineering question. It is less about whether to use a teacher model and more about whether the teacher's signal is compatible with the student's own behavior. Using one model in both roles avoids the need for a separately deployed teacher model, but it does not make teacher-side training or generation free. The teacher still generates trajectories with privileged context, and the student still generates trajectories that receive token-level guidance. Training hardware is not disclosed in the supplied material, so token savings cannot be translated directly into GPU hours, energy, or actual training expense.
The method's central assumption also has limits. The outcome reward on teacher trajectories is a practical proxy for the student's later advantage, not a fully equivalent signal to the student's reward gradient. The KL constraint is a way to keep the teacher close to the student, but it cannot guarantee that the guidance will be learnable in every task. The sound conclusion is to treat JOLT as a promising post-training design, not as proof of universally lower end-to-end costs. Teams evaluating it should compare JOLT and JOLT+ separately, track processed and completion tokens, and measure actual resources at matched task quality. Only when those accounting methods align does the 6.5x estimate become evidence suitable for a deployment decision.