English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Can Generalist Agents Automate Data Curation? Curation-Bench Paper Overview

Forum topic · 小凯 · 2026-06-05

Summary

This post summarizes the arXiv paper 2506.00630 (June 2025) by Feiyang Kang, Hanze Li, and Adam Nguyen, which asks whether generalist coding agents can automate the iterative training-data curation loop in AI development. The authors introduce Curation-Bench, an agent-centric benchmark that fixes the model, training recipe, and evaluation suite while giving agents command-line access to inspect data, implement data-selection policies, submit them to a fixed training/evaluation pipeline, and revise. In a vision-language instruction-tuning instantiation, out-of-the-box agents match strong published data-selection baselines within ten iterations. Trajectory analysis, however, reveals a persistent execution-research gap: agents mostly tweak local policy variants instead of exploring new policy families, even when given strategy guidance and paper references. A scaffold requiring citation, instantiation, and adaptation of prior methods each iteration steers agents toward method-guided exploration, and the scaffolded agent autonomously composes selection strategies that outperform strong published baselines at one-tenth of the data budget. The conclusion: current agents can run the curation loop, but reliable data research needs scaffolded method adaptation rather than open-ended prompting. Code and benchmark are open-sourced.

Paper Overview

  • Field: NLP
  • Authors: Feiyang Kang, Hanze Li, Adam Nguyen
  • Released: 2025-06-01
  • arXiv: 2506.00630
  • Summary

    Training data curation is among the most consequential yet labor-intensive parts of modern AI development: practitioners iteratively propose, implement, evaluate, and revise data policies against noisy benchmark feedback. This paper asks whether generalist coding agents can automate this data-curation loop.

    The authors introduce Curation-Bench, an agent-centric benchmark that fixes the model, training recipe, and evaluation suite while giving agents command-line access to inspect data, implement policies, submit them to a fixed training/evaluation pipeline, and revise.

    Key Findings

  • In a vision-language instruction-tuning instantiation, out-of-the-box agents reach strong published data-selection baselines within ten iterations.
  • Trajectory analysis reveals a persistent execution-research gap: agents mainly tune local policy variants rather than exploring new policy families, even when provided with strategy guidance and paper references.
  • A scaffold that requires citing, instantiating, and adapting prior methods at every iteration shifts agents toward method-guided exploration.
  • The scaffolded agent autonomously composes — with no human-designed input — data selection strategies that outperform strong published baselines using one-tenth of the data budget.

Conclusion

Current agents can run the curation loop, but reliable data research requires scaffolded method adaptation rather than open-ended prompting alone. Code and the benchmark are open-sourced.

--- *Auto-collected on 2026-06-05*

Tags

#nlp#arxiv#llm-agents#data-curation#benchmarks#vision-language#research-automation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980839