A new paper proposes a unified framework for generating high-quality training data for LLM agents, representing agentic data as a factorized object comprising environment specifications, task signals, interactions, and verifiers. The authors introduce the ACE (Accuracy-Complexity-diversity) lens, which formulates data generation as constrained distribution design: accuracy ensures execution-grounded consistency, complexity calibrates task difficulty to the learner’s current capability, and diversity controls coverage beyond surface-level variation. The work argues that improving agentic data quality is less about generating larger datasets and more about allocating valid, informative experiences as agents evolve.
