Perplexity and NVIDIA unveil JOLT for efficient LLM post-training

Researchers from Perplexity and NVIDIA have developed JOLT, a novel post-training method for language models that utilizes a single model acting as both teacher and student. This approach trains the teacher with privileged information while the student learns without it, but crucially, the teacher is also trained to ensure its feedback remains relevant. JOLT has demonstrated significant efficiency gains, matching GRPO's accuracy on the MATH dataset with 6.5 times fewer processed training tokens and achieving higher scores across various benchmarks like AppWorld and LiveCodeBench v6. AI

IMPACT This method could significantly reduce the computational cost of fine-tuning large language models, making advanced capabilities more accessible.

RANK_REASON The item describes a new method for language model post-training developed by researchers from Perplexity and NVIDIA, detailing its mechanics and performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on MarkTechPost →

• AppWorld
• DeepSeek-V2-Lite
• GRPO
• LiveCodeBench v6
• NVIDIA
• On-Policy Self-Distillation
• Perplexity
• Qwen3 0.6B
• Qwen3-30B-A3B-Instruct
• Qwen3-4B-Instruct
• Qwen3_8B

• model release
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →