What Makes Good Agentic Data? An ACE Framework for LLM Agents

Summary

The paper introduces a two-level framework for generating training data for LLM agents. It models experience through the environment, task, interaction, and optional verifier, then applies the ACE lens: execution-grounded accuracy, learner-relative complexity, and diversity. The authors argue that useful agentic data requires more than scale, and review methods for verification, difficulty calibration, and behavioral coverage.