返回全部动态
信任分层图书馆与多代理写作系统实现无漂移时间点研究报告
原标题：Reconcile Once, Write Anytime: A Trust-Tiered Librarian and a Multi-Agent Writer for Drift-Free, Point-in-Time Research
AI 摘要
该论文提出一个两层级代理系统，将维护的时间点知识库与报告写作分离，以解决长文研究报告中数值漂移、来源缺失和信任扁平化问题。系统在包含6130个来源的公共语料上评估，消除了6845个跨章节数字矛盾，并实现零前瞻违规。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
Reconcile Once, Write Anytime: A Trust-Tiered Librarian and a Multi-Agent Writer for Drift-Free, Point-in-Time ResearchThanks: *Corresponding author: peiyan@amazon.com. Abstract. Long-form research reports generated by large language models (LLMs) drift, contradict themselves, and lose provenance: the same metric appears with different values in different sections, numbers arrive without sources, and rumor is quoted as confidently as an audited filing. We present a deployment-oriented two-tier a
发布时间：2026-08-14 12:00
抓取时间：2026-08-14 12:39
来源机构：arXiv