Causality is Key for Interpretability Claims to Generalise
Mila - Qu\'ebec AI Institute \& Universit\'e de Montr\'eal；Boston University；Cold Spring Harbor Laboratory；Max-Planck-Institute for Intelligent Systems, ELLIS Institute T\"ubingen, University of T\"ubingen
针对大语言模型可解释性研究中普遍存在的泛化不足和因果解释缺乏证据的问题，本文强调因果推断在定义模型激活到不变高层结构的映射、所需数据或假设以及支持的推断方面的作用。通过 Pearl 的因果层级，本文澄清了可解释性研究能支持哪些结论。观察建立了模型行为与内部组件之间的关联，干预（如消融或激活修补）支持这些编辑如何影响行为指标（如平均词概率变化）的主张。然而，反事实主张——即在未观察干预下相同提示的模型输出——在缺乏受控监督的情况下难以验证。本文展示了因果表示学习（CRL）如何操作化这一层级，明确哪些变量可以从激活中恢复以及在何种假设下。这些内容共同推动了一种诊断框架，帮助从业者选择与主张相匹配的方法和评估，以确保研究结果的泛化。
英文摘要：
Interpretability research on large language models (LLMs) has yielded important insights into model behaviour, yet recurring pitfalls persist: findings that do not generalise, and causal interpretations that outrun the evidence. Our position is that causal inference specifies what constitutes a valid mapping from model activations to invariant high-level structures, the data or assumptions needed to achieve it, and the inferences it can support. Specifically, Pearl's causal hierarchy clarifies what an interpretability study can justify. Observations establish associations between model behaviour and internal components. Interventions (e.g., ablations or activation patching) support claims how these edits affect a behavioural metric (e.g., average change in token probabilities) over a set of prompts. However, counterfactual claims -- i.e., asking what the model output would have been for the same prompt under an unobserved intervention -- remain largely unverifiable without controlled supervision. We show how causal representation learning (CRL) operationalises this hierarchy, specifying which variables are recoverable from activations and under what assumptions. Together, these motivate a diagnostic framework that helps practitioners select methods and evaluations matching claims to evidence such that findings generalise.
Université Paris-Saclay（巴黎-萨克雷大学）；CentraleSupélec（中央理工-高等电力学院）；Saint-Gobain（圣戈班）；Université de Lorraine（洛林大学）；Sorbonne Université（索邦大学）；CEA（法国原子能和替代能源委员会）
Emerging Synergies in Causality and Deep Generative Models: A Survey
University of New South Wales（新南威尔士大学）；Carnegie Mellon University（卡内基梅隆大学）；Mohamed bin Zayed University of Artificial Intelligence（穆罕默德·本·扎耶德人工智能大学）；University of California, San Diego（加利福尼亚大学圣迭戈分校）；CSIRO’s Data61（澳大利亚联邦科学与工业研究组织Data61研究院）
Causality and Interpretability for Electrical Distribution System faults
ABB Ability Innovation Center（ABB Ability创新中心）；Department of Electronics and Communication Engineering（电子与通信工程系）；IIT Bhubaneswar（比哈尔邦理工学院）；Department of Artificial Intelligence（人工智能系）；IIT Hyderabad（海得拉巴理工学院）；Department of Computer Science and Engineering（计算机科学与工程系）
SplInterp: Improving our Understanding and Training of Sparse Autoencoders
School of Mathematics University of Birmingham（数学学院英国伯明翰大学）；Ideami Studios（Ideami工作室）；Department of Mathematics and Statistics University of Limerick（数学与统计学学院英国利默里克大学）；XRAI Inc.（XRAI公司）；Department of Computer Science Brown University（计算机科学系布朗大学）
ICST-DNET: An Interpretable Causal Spatio-Temporal Diffusion Network for Traffic Speed Prediction
College of Computer and Software, Hohai University（河海大学计算机与软件学院）；School of Computer Science and Engineering, Nanyang Technological University（南洋理工大学计算机科学与工程学院）；College of Internet of Things, Nanjing University of Posts and Telecommunications（南京邮电大学物联网学院）
Towards Trustworthy and Aligned Machine Learning: A Data-centric Survey with Causality Perspectives
Bundelkhand Institute of Engineering and Technology（本德尔坎德工程技术学院）；Nanyang Technological University（南洋理工大学）；University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）
Federated Causality Learning with Explainable Adaptive Optimization
School of Software, Shandong University（山东大学软件学院）；SDU-NTU Joint Centre for AI Research, Shandong University（山东大学-新加坡南洋理工大学人工智能联合研究中心）；Department of Mathematics, National University of Singapore（新加坡国立大学数学系）；Departiment of Computer Science, George Mason University（乔治梅森大学计算机系）；School of Control Science and Engineering, Shandong University（山东大学控制科学与工程学院）
针对隐私保护下分散异构数据的因果发现问题，提出 FedCausal 联邦策略，通过可解释的自适应优化统一局部与全局 DAG 学习，在 non-iid 合成及真实数据上表现优越。
Causal Inference via Nonlinear Variable Decorrelation for Healthcare Applications
University of Massachusetts Amherst（马萨诸塞大学阿默斯特分校）；University of Rochester（罗切斯特大学）；Shanghai Maritime University（上海海事大学）；University of Rochester Medical Center（罗切斯特大学医学中心）
Causality Is Key to Understand and Balance Multiple Goals in Trustworthy ML and Foundation Models
CISPA Helmholtz Center for Information Security（CISPA海德堡信息安全中心）；Max Planck Institute for Intelligent Systems, Tübingen（马克斯·普朗克智能系统研究所（图宾根））；Google Research（谷歌研究）；ETH Zürich（苏黎世联邦理工学院）；University of Toronto（多伦多大学）
Predicting Subway Passenger Flows under Incident Situation with Causality
Key Laboratory of Road and Traffic Engineering, Ministry of Education at Tongji University（同济大学交通工程教育重点实验室）；Urban Mobility Institute, Tongji University（同济大学城市交通研究院）
Semi-Supervised Learning, Causality and the Conditional Cluster Assumption
Max Planck Institute for Intelligent Systems（马克斯·普朗克智能系统研究所）；University of Cambridge（剑桥大学）；Delft University of Technology（代尔夫特理工大学）；University of Copenhagen（哥本哈根大学）
Tsinghua University（清华大学）；Guangdong University of Technology（广东工业大学）；Peng Cheng Laboratory（鹏城实验室）；Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；Shantou University（汕头大学）
Interpretability and causal discovery of the machine learning models to predict the production of CBM wells after hydraulic fracturing
Southwest Petroleum University（西南石油大学）；Institute for Artificial Intelligence, Southwest Petroleum University（西南石油大学人工智能研究院）；State Key Laboratory of Oil and Gas Reservoir Geology and Exploitation（油气藏地质及开发工程国家重点实验室）
CRTRE: Causal Rule Generation with Target Trial Emulation Framework
University of Massachusetts Amherst（马萨诸塞大学阿默斯特分校）；University of Rochester（罗切斯特大学）；Shanghai Maritime University（上海海事大学）；University of Massachusetts Lowell（马萨诸塞大学洛厄尔分校）
Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond
Technion - Israel Institute of Technology（以色列理工学院）；Princeton University（普林斯顿大学）；Williams College（威廉姆斯学院）；University of Wisconsin - Madison（威斯康星大学麦迪逊分校）；Microsoft（微软）；Columbia University（哥伦比亚大学）；Northwestern University（西北大学）；Google Research（谷歌研究院）；Stanford University（斯坦福大学）；University of California San Diego（加州大学圣地亚哥分校）；University of Chicago（芝加哥大学）；Georgia Tech（佐治亚理工学院）
