ACE lens reframes what counts as good agent training data

LLM agents are increasingly trained on data those same models help generate, and a new arxiv paper from Xingshan Zeng and thirteen coauthors argues the field has no shared way to say what "good" agentic data even means. Their proposal is the ACE lens: Accuracy, Complexity, and divErsity.

The paper first represents any piece of agentic data as a four-part object: an environment specification, a task signal, an interaction realization, and an optional verifier. On top of that, it treats generation as "constrained distribution design," with accuracy fixing the feasible support of grounded data, complexity placing "learning mass relative to the capability of a declared learner," and diversity controlling coverage and redundancy.

The headline argument is blunt. "The central challenge is not simply to generate more data, but to continually allocate valid, informative, and non-redundant experience as agents and environments evolve," the authors write. They report the literature is already drifting that way, toward "execution-grounded accuracy, learner-relative complexity, and diversity beyond surface variation or dataset size."

The abstract advertises no benchmark, no per-dataset ACE scores, and no metric formulas for the three letters. It is a map of a fragmented literature, not a scoreboard.