New SAZO method accelerates LLM fine-tuning with adaptive sample selection

Researchers have introduced Sample-Adaptive Zeroth-Order Optimization (SAZO), a novel method for fine-tuning large language models that improves upon existing Zeroth-order (ZO) optimization techniques. Unlike traditional ZO methods that sample training examples uniformly, SAZO adaptively learns sample importance online, allocating more queries to critical examples to accelerate convergence. This approach extracts gradient-norm information to estimate sample importance without needing true gradients and maintains per-example moving averages to track evolving importance. Experiments show SAZO achieves faster convergence and improved downstream performance across various model families. AI

IMPACT SAZO's adaptive sampling could significantly reduce the computational cost and time required for fine-tuning LLMs, making advanced model customization more accessible.

RANK_REASON The cluster describes a new method presented in an academic paper on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Hugging Face
• large language models
• MeZO
• Sample-Adaptive Zeroth-Order Optimization
• Zeroth-order optimization

• paper
• infra

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