Autodata: An agentic data scientist to create high quality synthetic data

The Autodata Framework

Autodata is a systematic approach to synthetic data generation that moves beyond simple prompting. It treats the creation of training data as an iterative, agentic process. An autonomous agent acts as a data scientist, performing a cycle of data generation, qualitative inspection, and quantitative performance evaluation. By analyzing how different models (a 'weak' solver and a 'strong' solver) perform on generated tasks, the agent learns to refine its data-generation recipe to produce examples that are specifically calibrated to improve model reasoning.

Agentic Self-Instruct

The authors implement this framework through 'Agentic Self-Instruct.' In this setup, a main orchestrator agent manages four subagents: a challenger (which creates the data), a weak solver, a strong solver, and a verifier/judge. The system aims to generate examples where the strong solver succeeds but the weak solver struggles, creating a 'learning gap' that is ideal for training the weaker model. If the generated data fails to meet these criteria, the judge provides feedback, and the challenger iterates until an acceptable example is produced.

Performance Gains

In experiments involving computer science research papers and legal reasoning, the Autodata approach consistently outperformed standard CoT (Chain-of-Thought) Self-Instruct methods. For computer science tasks, the agentic loop produced questions that were significantly more challenging, leading to better performance when used to train a 4B parameter model via Reinforcement Learning (RL). In the legal domain, where standard synthetic data was often too difficult for the model to learn from, the agentic loop successfully adjusted the difficulty to create more suitable training material, demonstrating the method's flexibility.

As AI models become more capable, the demand for high-quality, challenging training data grows. Autodata provides a scalable way to convert increased inference compute into higher-quality training data. By automating the curation and refinement process, this method allows researchers to generate benchmarks and training sets that specifically target model weaknesses, potentially accelerating progress on complex reasoning tasks where human-generated data is scarce or expensive to produce.