Perplexity's JOLT matches GRPO with 6.5x fewer training tokens, JOLT+ wins on all five benchmarks
perplexity_ai · x · 2026-10-22
Perplexity introduces JOLT, a post-training method where a single model acts as both teacher and student, with the teacher trained to give feedback the student can actually use.
Key results
- On MATH, JOLT matches GRPO's accuracy (63.44%) with 6.5x fewer processed training tokens (1.30M vs 8.40M), using Qwen3-0.6B as base.
- JOLT+ beats GRPO across all five benchmarks spanning math, coding, tool use, and terminal tasks: GSM8K +1.5, MATH +4.2, LCBv6 +3.0, and AppWorld +9.9 points (72.09% vs 62.20%).
How it works
- RL methods like GRPO give one reward per attempt, showing nothing about which steps helped.
- JOLT uses self-distillation: the same shared-weight model plays teacher (with privileged information like a verified solution) and student (without it). Student generations provide the distillation signal; teacher generations scored without the privileged info incur a KL penalty when they diverge from the student, preventing the teacher from exploiting shortcuts unavailable to the student.
- The teacher is also trained on outcome rewards from its own attempts.
Paper and code are available.
Related event: Perplexity releases JOLT: 6.5x fewer training tokens than GRPO(2 posts)→
More from Models
- Claude Duet is built on Opus 5.5, inheriting memory, skills and apps — claudeai · 2026-10-22
- Feedback Bench ranks personal AI agents by real user feedback, Meta Muse tops at 70.5 — aakashgupta · 2026-10-22
- 218K user posts ranked: Meta's Muse tops personal AI agents at 70.5 — aakashgupta · 2026-10-22
- Why do frontier models keep asking to be 'the guy with a ruler and a ledger'? — repligate · 2026-10-22
- Flint model weights hit HuggingFace, runs on DGX Spark, RTX GPUs and M4 Pro — testingcatalog · 2026-10-22
- Claude Sonnet 5.5 at max reasoning writes worse code than xhigh while burning most tokens — DeepLearningAI · 2026-10-22