Reproduce “CurateEvo: Data-Curation Evolving for Agentic Post-Training” on OpenResearch

About this paper

Dingzirui Wang, Xuanliang Zhang, Keyan Xu, Qingfu Zhu, Wanxiang Che

CurateEvo adapts data curation for tool-using language-model agents by rewriting executable rules from failed development trajectories. The rules produce supervised fine-tuning, reinforcement-learning, and inference-time memory data while targeting failures and pruning redundant turns. On three held-out benchmarks, it leads prior methods in labeled and wild settings, with equal-weight gains of 3.2 and 2.7 points, though comparisons are not uniformly model-matched.

Key insights

• Failure-driven adaptation, which involves diagnosing and targeting specific agent shortcomings, is the primary driver of performance improvement in agentic post-training data curation.
• Jointly optimizing for both data effectiveness (improving agent capabilities) and efficiency (reducing training cost) is crucial for achieving optimal agent performance and practical feasibility.
• The framework highlights the importance of comprehensive data curation operations—augmentation, filtering, and refinement—over merely augmenting data, to create higher-quality training signals.

Summary by alphaXiv.