Can Generalist Agents Automate Data Curation?
This work addresses the labor-intensive data curation process in AI development, demonstrating that current agents can automate the loop but require scaffolded method adaptation for reliable data research.
The paper introduces Curation-Bench, a benchmark for evaluating generalist coding agents on automating the data curation loop. Results show that out-of-the-box agents match strong baselines within ten iterations, and a scaffolded agent autonomously composes a policy that outperforms published baselines at one-tenth the data budget.
Curating training data 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. We ask whether generalist coding agents can automate this data-curation loop. We 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. In a vision-language instruction-tuning instantiation, out-of-the-box agents reach strong published data-selection baselines within ten iterations. However, trajectory analysis reveals a persistent *execution-research gap*: agents mainly tune local policy variants rather than explore new policy families, even when given strategy guides and paper references. Scaffolds requiring each iteration to cite, instantiate, and adapt a prior method shift agents toward method-guided exploration. The scaffolded agent autonomously composes -- without human design input -- a data-selection policy that outperforms strong published baselines at one-tenth their data budget. Overall, current agents can run the curation loop, but reliable data research requires scaffolded method adaptation, not open-ended prompting alone. Code and benchmark are open-sourced.