Perplexity's JOLT Beats GRPO on All Five Benchmarks

Perplexity introduced JOLT, a post-training method where one model plays both teacher and student, using per-token feedback instead of scalar rewards. JOLT needs only 1/6.5 of GRPO's training tokens on MATH, and JOLT+ beats GRPO across all five benchmarks.

2026-10-22 ~ 2026-10-22 · 4 related posts

• Perplexity's JOLT matches GRPO with 6.5x fewer training tokens, JOLT+ wins on all five benchmarks — perplexity_ai · 2026-10-22
• GRPO gives one reward per attempt; JOLT adds per-token teacher feedback instead — perplexity_ai · 2026-10-22
• JOLT trains the teacher too, with penalties when its predictions diverge from the student — perplexity_ai · 2026-10-22
• JOLT+ beats GRPO on all five benchmarks, up 9.9 points on AppWorld — perplexity_ai · 2026-10-22