AIAny
Icon for item

What Makes Good Agentic Data? An ACE Lens on Data Generation for LLM Agents

Analyzes how to generate useful interaction data for LLM agents and proposes the ACE lens — Accuracy, Complexity, divErsity — while factorizing agentic data as (E, q, τ, v). Surveys verification, difficulty calibration, and coverage strategies and outlines implications for training and benchmarks.

Introduction

Agentic systems increasingly learn from synthetic interaction traces rather than purely human-curated corpora. The core insight of this paper is that quality for agentic data is not just quantity: useful experience must be simultaneously valid (grounded), appropriately difficult relative to the learner, and non-redundant. The authors package this intuition into a concise analytical lens — Accuracy, Complexity, divErsity (ACE) — and a factorized representation of an agentic datum as (environment spec E, task signal q, interaction τ, optional verifier v).

Key Findings
  • Factorized representation clarifies roles: treating data as (E, q, τ, v) separates environment grounding from task design, interaction realization, and verification responsibilities, which helps compare generation pipelines.
  • Accuracy means execution-grounded and internally consistent samples; the paper argues verification should move from surface checks to execution or simulator grounding to avoid spurious signals.
  • Complexity should be allocated relative to a declared learner and execution setup: samples near the learner’s capability boundary (informative failures and marginal successes) provide disproportionate learning signal.
  • divErsity covers more than surface variability or dataset size; it includes structural coverage (different API topologies, dependency depths) and redundancy control to avoid wasted labeling and overfitting.
  • Across prior work, the field is shifting toward execution-grounded accuracy, learner-relative difficulty calibration, and diversity metrics that prioritize behavioral coverage and informative edge cases.
Who it's for & tradeoffs

Great fit if you design or curate synthetic interaction datasets for LLM agents, build simulators/API-based trainers, or evaluate agent benchmarks — the ACE lens gives concrete criteria to decide what samples to synthesize, verify, and emphasize. Look elsewhere if you only need short how‑to recipes: this paper is analytical and comparative rather than a single turnkey dataset or simulator. A key tradeoff highlighted is cost vs. fidelity: execution-grounded verification and learner-relative resampling improve utility but increase generation and validation overhead.

Where it fits

This paper sits between methods papers that propose particular synthetic pipelines and surveying works that list datasets: it provides a unifying vocabulary and prescriptive priorities that can guide both dataset builders (what to generate) and evaluators (what to measure).

Practical implication

Adopt ACE when designing generation loops: (1) ensure environmental grounding or a rigorous simulator for Accuracy; (2) tune sampling toward capability-boundary tasks for Complexity; (3) measure and control structural coverage, not just token-level diversity, for divErsity. These shifts help convert extra compute into higher-quality training and evaluation data as agents and environments evolve.

Information

  • Websitearxiv.org
  • AuthorsXingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang
  • Published date2026/08/27

More Items

Provides a real-scale 3D Hong Kong sandbox to evaluate whether multimodal LLM agents can turn local street-view perception into sustained spatial action, supporting closed-loop first-person interaction, an interactive map, and controlled tests of grounding, long-range navigation, and robustness.

Trains compact conversational agents to adapt at runtime to changing 'Harness' configurations (Skills, Hooks, prompts, tools) using Harness-Aware Training (HAT): Harness-State Augmentation, on-policy distillation, and RL to preserve generality while meeting low-latency deployment constraints.

A test-time method that adapts LLMs without labels by distilling rollouts that agree with majority pseudo-labels and penalizing disagreeing rollouts via grouped RL, improving robustness under frequent pseudo-label errors.