LLM-AutoDP: The Framework That Lets an LLM Agent Design Its Own Training Data Pipeline
The most expensive step in fine-tuning a domain-specific LLM is not training. It is the human expert sitting in front of medical records deciding which samples to keep, which to clean, and in what order to apply the cleaning operations. LLM-AutoDP (Ant Group, VLDB 2026) eliminates that expert by using an LLM agent to iteratively generate, evaluate, and refine data processing strategies, with three acceleration techniques that cut total search time by up to 10x and a privacy architecture that means no human ever sees the raw data.
