New curriculum learning method efficiently distills CoT reasoning into smaller models

Researchers have developed a novel three-stage curriculum learning framework to distill Chain-of-Thought (CoT) reasoning from large language models into smaller, more efficient models. This method employs structure-aware masking and Group Relative Policy Optimization (GRPO) to progressively enhance the student model's ability to reproduce teacher rationales without excessive verbosity. Experiments on the GSM8K benchmark showed that Qwen2.5-3B-Base, using this distillation technique, achieved an 11.29% accuracy increase while reducing output length by 27.4%, outperforming existing distillation methods. AI

IMPACT This technique could enable more efficient deployment of powerful reasoning capabilities in resource-constrained environments.

RANK_REASON The cluster contains a research paper detailing a new method for model distillation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Bowen Yu
• Chain-of-Thought
• Group Relative Policy Optimization
• GSM8K
• Hugging Face
• Qwen2.5-3B-Base

• paper
• model release

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