The ACE lens reframes agentic training data as a factorized object with an optional verifier

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 27 August, this survey argues the field's domain-by-domain organization hides shared generation mechanisms and conflates candidate construction with verification and selection. It represents agentic data as a factorized object (E, q, tau, v): environment specification, task signal, interaction realization, and optional verifier, then organizes generation paradigms by primary anchor and dependency structure. The second level treats generation as constrained distribution design under Accuracy, Complexity and divErsity, where accuracy establishes the feasible support of grounded data, complexity places learning mass relative to a declared learner and execution configuration, and diversity controls coverage and redundancy.