Summary
- A comprehensive survey paper titled Data Optimization for LLMs by Ou Wu has been accepted for publication in ACM TKDD.
- The study introduces a framework for LLM data optimization, addressing data challenges in training and inference.
- It outlines data perception dimensions including quality, difficulty, diversity, and redundancy.
- Optimization techniques are categorized into selection, augmentation, reweighting, and other methods, alongside future research directions.
Why it matters
Data quality and optimization are critical bottlenecks in modern AI development. This survey synthesizes disparate techniques into a unified framework, helping practitioners improve data efficiency and model performance.