Workflow is All You Need: Escaping the "Statistical Smoothing Trap" via High-Entropy Information Foraging and Adversarial Pacing
流程即一切:通过高熵信息觅食和对抗性节奏控制摆脱'统计平滑陷阱
机构 * Zhongjie Jiang(江宗杰)
专题命中 工作流自动化 :workflow(title,abstract);agentic(abstract,comments);planning(abstract);分类 cs.AI、cs.CL
AI总结 本研究提出DeepNews框架,通过高熵信息觅食和对抗性节奏控制,解决LLMs在长文本生成中的幻觉、逻辑一致性和个性化表达难题,实验显示其在金融报道中表现优异。
Comments 22 pages, 8 figures. Includes an ecological validity blind test where the Agentic Workflow achieved a 25% acceptance rate in top-tier media, decisively outperforming the SOTA Zero-shot baseline (0%). Features the DNFO-v5 ontology