What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents
The paper argues that agent training data should be judged by validity, difficulty, and coverage, not volume.
It defines agentic data as environment, task, interaction, and optional verifier, then uses that structure to compare generation methods. Its ACE lens treats accuracy as the grounded support, complexity as learner-relative difficulty, and diversity as coverage without redundancy. The authors say the field is moving toward execution-grounded checks, calibrated difficulty, and diversity beyond surface variation. HF Daily Papers' note
It defines agentic data as environment, task, interaction, and optional verifier, then uses that structure to compare generation methods. Its ACE lens treats accuracy as the grounded support, complexity as learner-relative difficulty, and diversity as coverage without redundancy. The authors say the field is moving toward execution-grounded checks, calibrated difficulty, and diversity beyond surface variation. HF Daily Papers' note
score 4