Reproduce “CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yanbei Jiang, Chao Lei, Rena Gao, Xiaochen Zhao, Yuchen Ge, Haohang Liu, Runze Nie, Qitian Chen and 12 more
CARE trains customer-service models from verified production failures. It attributes failures, truncates examples at the first model error, and uses a frozen self-teacher with privileged hindsight unavailable to the student. After five iterations, V5 scored 96.07 on the combined offline measure, nearly matching GPT-5.4’s 96.10, while online satisfaction was statistically at parity; testing covered only latest SOP and an unvalidated simulator.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.