Earlier research has shown that metaphors influence human decision-making, raising the question of whether metaphors also influence large language models (LLMs)' reasoning pathways, given that their training data contain a large number of metaphors. In this work, we investigate the problem in the scope of the emergent misalignment problem, where LLMs can generalize patterns learned from misaligned content in one domain to another domain. We find strong evidence that metaphors in training data contribute to cross-domain misalignment in LLMs' reasoning outputs. With metaphor-based interventions during continued pre-training and fine-tuning for inducing misalignment, models exhibit significantly different degrees of emergent cross-domain misalignment. We also observe similar effects in re-alignment settings. As we further investigate this phenomenon, we find that metaphors are linked to the activation of latent features in large reasoning models. By monitoring these latent features, we design a detector that predicts misaligned content with high accuracy.
Towards Multimodal Metaphor Understanding: A Chinese Dataset and Model for Metaphor Mapping Identification
Dalian University of Technology（大连理工大学）；School of Foreign Languages（外国语学院）；School of Software（软件学院）；University of Macau（澳门大学）；Faculty of Business Administration（工商管理学院）；Kansai University（关西大学）；Faculty of Business and Commerce（商学部）；University of New South Wales（新南威尔士大学）；School of Minerals and Energy Resources Engineering（矿产与能源资源工程学院）；RMIT University（皇家墨尔本理工大学）；Centre for Educational Innovation and Quality（教育创新与质量中心）；School of Computing Technologies（计算技术学院）
Elucidating STEM Concepts through Generative AI: A Multi-modal Exploration of Analogical Reasoning
Carnegie Learning（卡内基学习公司）；University of Sheffield（谢菲尔德大学）；Microsoft Research（微软研究院）；University of Maryland, College Park（马里兰大学帕克分校）；University of Georgia（佐治亚大学）；New York University（纽约大学）；Carnegie Mellon University（卡内基梅隆大学）
Dutch Metaphor Extraction from Cancer Patients' Interviews and Forum Data using LLMs and Human in the Loop
Biomedical Data Sciences, Leiden University Medical Center（莱顿大学医学中心生物医学数据科学）；Leiden Institute of Advanced Computer Science (LIACS), Leiden University（莱顿大学先进计算机科学研究所）；Department of Radiation Oncology (MAASTRO), GROW Research Institute for Oncology（肿瘤学研究与生殖学研究所放射肿瘤科）
Narrative Paradoxes in LLM Inference: Shaping Reasoning Trajectories toward Misalignment
School of Electronic, Electrical and Communication Engineering, UCAS（中国科学院大学电子电气与通信工程学院）；School of Computer Science and Technology, UCAS（中国科学院大学计算机科学与技术学院）；Nanyang Technological University, Singapore（新加坡南洋理工大学）；State Key Laboratory of AI Safety, Institute of Computing Technology, CAS（中国科学院计算技术研究所人工智能安全重点实验室）；Beijing Academy of Artificial Intelligence（北京智源人工智能研究院）；Alibaba Group（阿里巴巴集团）
"Dark Triad" Model Organisms of Misalignment: Narrow Fine-Tuning Mirrors Human Antisocial Behavior
Brain & Creativity Institute, University of Southern California（大脑与创造力研究所，南加州大学）；Department of Psychology, University of Southern California（心理学系，南加州大学）；Interchange Forum for Reflecting on Intelligent Systems, University of Stuttgart（智能系统反思交流论坛，斯图加特大学）
Reasoning in Conversation: Solving Subjective Tasks through Dialogue Simulation for Large Language Models
Department of Computer Science and Technology, Tsinghua University（清华大学计算机科学与技术系）；Institute for AI Industry Research (AIR), Tsinghua University（清华大学人工智能产业研究院）
Preemptive Detection and Correction of Misaligned Actions in LLM Agents
Ubiquitous Knowledge Processing Lab, Department of Computer Science, TU Darmstadt（图宾根大学计算机科学系）；National Research Center for Applied Cybersecurity ATHENE, Germany（德国应用网络安全国家研究中心ATHENE）；Department of Electrical and Computer Engineering & Ingenuity Labs Research Institute, Queen’s University, Canada（加拿大女王大学电气与计算机工程系及创新实验室研究院）
DoReMi: Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment
Institute for Interdisciplinary Information Sciences, Tsinghua University（清华大学交叉信息研究院）；Weiyang College, Tsinghua University（清华大学未央学院）；Shanghai Qi Zhi Institute（上海期智研究院）
From Consistency to Complementarity: Aligned and Disentangled Multi-modal Learning for Time Series Understanding and Reasoning
The Hong Kong University of Science（香港科学与技术大学）；Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；University of Chinese Academy of Sciences（中国科学院大学）
