LLM-AutoDP: Automatic Data Processing via LLM Agents for Model Fine-tuning

Researchers introduce LLM-AutoDP, a framework that uses large language models as autonomous agents to automatically optimize data processing strategies for fine-tuning without human intervention or direct data exposure. The system achieves over 80% win rates against baseline models and reduces search time by up to 10x through novel acceleration techniques, addressing critical challenges in domain-specific model training and data privacy.

LLM-AutoDP represents a significant advancement in automating machine learning workflows by eliminating manual trial-and-error processes that typically consume substantial resources. The framework addresses a persistent bottleneck in AI development: preparing high-quality training data for domain-specific applications. Traditional data processing requires extensive human review and iterative refinement, creating both operational costs and security vulnerabilities when sensitive information is involved. By leveraging LLMs as autonomous agents, the system generates and evaluates processing strategies in-context, enabling convergence toward optimal pipelines without exposing raw data to human reviewers.

The technical innovations underlying LLM-AutoDP reflect broader trends in AI automation and privacy-preserving machine learning. Distribution Preserving Sampling, Processing Target Selection, and the Cache-and-Reuse Mechanism collectively reduce computational overhead while maintaining data integrity. These optimizations are particularly valuable for healthcare, financial services, and other regulated sectors where privacy constraints limit manual intervention. The 65% win rate against competing LLM-based AutoML baselines indicates meaningful performance improvements over contemporary approaches.

The framework's implications extend across multiple stakeholder groups. For organizations deploying specialized LLMs, automated data processing reduces time-to-market and operational costs while mitigating privacy risks. For AI researchers, the agent-based optimization approach demonstrates how LLMs can effectively manage complex hyperparameter searches and pipeline design. The 10x acceleration in search time particularly benefits resource-constrained environments.

Future developments likely involve scaling the framework across different data modalities and domain-specific challenges. Integration with existing MLOps platforms and evaluation on larger industrial datasets will determine practical adoption rates. The intersection of autonomous optimization and privacy preservation positions this work at the frontier of enterprise AI deployment.

• →LLM-AutoDP automates data processing strategy generation without requiring human access to sensitive data, addressing privacy concerns in regulated industries.
• →Models trained on LLM-AutoDP-processed data achieve over 80% win rates against unprocessed baselines and 65% win rates against competing AutoML approaches.
• →Three key acceleration techniques reduce total search time by up to 10x while maintaining data integrity and processing quality.
• →The framework leverages in-context learning to iteratively refine processing strategies using feedback signals rather than manual trial-and-error.
• →The approach demonstrates broad applicability across domain-specific fine-tuning scenarios, particularly in healthcare and privacy-critical applications.