Tool-using language agents are increasingly deployed in regulated customer-service workflows, where getting the final answer right is not enough. According to a new paper on arXiv, agents must also provide disclosures, obtain consent, and follow required procedures. The authors introduce CXBench, a benchmark designed to measure this kind of process compliance.

CXBench targets scenarios where a correct final state is insufficient. The benchmark evaluates whether agents adhere to the steps mandated by regulations, rather than only checking whether the outcome matches an expected result. The abstract notes that existing approaches fall short, though the full details of the benchmark construction and results are not included in the available abstract.

Because the source is a single arXiv announcement, this article is based solely on the abstract. The paper's specific tasks, evaluation metrics, and findings are not described in the provided text.