DataCOPE: Unsupervised Skill Discovery for Agentic Data Analysis

arXiv·medium signal

DataCOPE is a verifier-guided framework that discovers reusable data-analysis skills from unlabeled exploration alone, injecting procedural knowledge at inference time without updating model weights. It addresses the problem that reliable supervision for data-analytic agents is expensive and success criteria vary across formats. Directly applicable to builders who want self-improving analytic agents without fine-tuning.