HumorPlanSearch:面向上下文 AI 幽默的结构化规划与 HuCoT
计算机视觉与模式识别
2026-03-03 v2
摘要
大型语言模型(LLMs)生成的自动化幽默往往产生通用、重复或脱离语境的笑话,因为幽默深植于情境之中,取决于听众的文化背景、心态和即时上下文。我们引入HumorPlanSearch,一个模块化管道,通过以下方式显式建模情境:(1) Plan-Search for diverse, topic-tailored strategies;(2) Humor Chain-of-Thought (HuCoT) templates capturing cultural and stylistic reasoning;(3) a Knowledge Graph to retrieve and adapt high-performing historical strategies;(4) novelty filtering via semantic embeddings;and (5) an iterative judge-driven revision loop。为评估情境敏感性和幽默质量,我们提出了幽默生成分数(HGS),该分数融合了直接评分、多人物反馈、成对获胜率和话题相关性。在针对九个话题、来自13名人类评判官的反馈实验中,我们的完整管道(KG + Revision)相对于强基线提升了15.4%的平均HGS(p < 0.05)。通过在策略规划到多信号评估的每个阶段都突出显示情境,HumorPlanSearch推动了AI驱动的幽默向更具连贯性、适应性和文化敏感性的喜剧发展。
引用
@article{arxiv.2508.11428,
title = {ImagiDrive: A Unified Imagination-and-Planning Framework for Autonomous Driving},
author = {Jingyu Li and Bozhou Zhang and Xin Jin and Jiankang Deng and Xiatian Zhu and Li Zhang},
journal= {arXiv preprint arXiv:2508.11428},
year = {2026}
}
备注
Accepted for publication in 2026 IEEE International Conference on Robotics and Automation (ICRA)