AI News Feed
Market watch
TECNO Releases Film on '100 Portraits of Becoming' With Angélica DassIBM releases Granite Speech 5.0 2B and forecasting ensemble; China Telecom open-sources intent modelPerplexity and NVIDIA Researchers Introduce JOLT Post-Training MethodSeoul stocks slide 3.52% as AI investment concerns spur chip profit-takingOpenAI adds ChatGPT sidebar and cross-tab help to ChromeBOK Sees Chip Export Boom Lifting Private Consumption; Lee Urges Balanced Growth
TECNO Releases Film on '100 Portraits of Becoming' With Angélica DassIBM releases Granite Speech 5.0 2B and forecasting ensemble; China Telecom open-sources intent modelPerplexity and NVIDIA Researchers Introduce JOLT Post-Training MethodSeoul stocks slide 3.52% as AI investment concerns spur chip profit-takingOpenAI adds ChatGPT sidebar and cross-tab help to ChromeBOK Sees Chip Export Boom Lifting Private Consumption; Lee Urges Balanced Growth
Large Language Models

Perplexity and NVIDIA Researchers Introduce JOLT Post-Training Method

Perplexity and NVIDIA researchers introduce JOLT, a method training one model as teacher and student, cutting MATH training tokens 6.5x versus GRPO.

AI News Feed · 2026-10-22 16:47

JOLT builds on On-Policy Self-Distillation, or OPSD. In JOLT, the teacher receives privileged information, such as a verified solution, while the student sees only the task. The key change is that the teacher is also trained, receiving outcome rewards plus a KL penalty that keeps it close to the student; the student learns from the teacher's token-level distillation signal on its own attempts. JOLT+ adds the student's own outcome reward to the student update. The researchers say a teacher holding the answer can take shortcuts the student cannot follow, and they reported that OPSD's teacher-student disagreement grew during MATH training.

Tests covered five settings: GSM8K with DeepSeek-V2-Lite; MATH with Qwen3-0.6B; LiveCodeBench v6 with Qwen3-8B on 131 problems; AppWorld with Qwen3-4B-Instruct; and Terminal-Bench 2.1 with Qwen3-30B-A3B-Instruct trained on TMax. Each setting used a different type of privileged information, including verified answers, worked reference solutions, a passing sibling solution or execution feedback, summaries of past attempts and private verifier files.

JOLT+ posted the highest mean score in every setting tested against GRPO, OPSD and π-Distill, according to Perplexity's own runs reported in the paper. Against GRPO, it gained 1.5 points on GSM8K, 4.2 on MATH, 3.0 on LiveCodeBench v6 and 9.9 on AppWorld. On Terminal-Bench 2.1, JOLT+ reached 10.96% pass@1 and 34.83% pass@32, compared with 8.25% and 30.34% for GRPO. Plain JOLT, without student reward, beat GRPO on MATH, 72.14% versus 68.36%. JOLT without student reward scored 51.92% on AppWorld, below GRPO's 62.20%, while JOLT+ reached 72.09% there.

On MATH, the researchers fit accuracy curves against training-token budgets. At 63.44% accuracy, GRPO's checkpoint used 8.40 million processed training tokens, while JOLT's fitted curve needed 1.30 million, a 6.5x reduction. Counting completion tokens only widened the gap to about 13x. Across 55% to 65% accuracy targets, savings ranged from 4.3x to 7.4x on processed tokens. JOLT also plateaued near 72%, against about 70% for GRPO and 62% for OPSD.

The report did not disclose training hardware. It listed the code as not yet released. The tested base models included Qwen3-0.6B, Qwen3-4B-Instruct, Qwen3-8B, Qwen3-30B-A3B-Instruct and DeepSeek-V2-Lite.