Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias often leads to unexpected model failures and hurts performance, robustness, and reliability across various applications. Our mechanistic analysis attributes the position bias to two components employed in nearly all state-of-the-art LMs: causal attention and relative positional encodings. Based on the analyses, we propose to eliminate position bias (e.g., different retrieved documents' orders in QA affect performance) with a training-free zero-shot approach. Our method changes the causal attention to bidirectional attention between documents and utilizes model attention values to decide the relative orders of documents instead of using the order provided in input prompts, therefore enabling Position-INvariant inferencE (PINE) at the document level. By eliminating position bias, models achieve better performance and reliability in downstream tasks, including LM-as-a-judge, retrieval-augmented QA, molecule generation, and math reasoning. Notably, PINE is especially useful when adapting LMs for evaluating reasoning pairs: it consistently provides 8 to 10 percentage points performance gains, making Llama-3-70B-Instruct perform even better than GPT-4-0125-preview and GPT-4o-2024-08-06 on the RewardBench reasoning set.
Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias
AMAP, Alibaba Group（阿里巴巴集团AMAP实验室）；Mohamed bin Zayed University of Artificial Intelligence（穆罕默德·本·扎耶德人工智能大学）；University of Southern California（南加州大学）；University of Michigan, Ann Arbor（密歇根大学安娜堡分校）
Analysing Moral Bias in Finetuned LLMs through Mechanistic Interpretability
Department of Computer Science and Engineering, University of Bologna, Bologna, Italy（计算机科学与工程系，博洛尼亚大学，意大利博洛尼亚）；Center for Studies and Research in Cognitive Neuroscience, Department of Psychology, University of Bologna, Cesena, Italy（认知神经科学研究中心，心理学系，博洛尼亚大学，意大利塞森纳）
Mechanistic Diagnostics of Spatial Lexical Bias in Multimodal Large Language Model Spatial Reasoning
Kyoto University（京都大学）；NII LLMC（日本国立信息与通信技术研究所语言模型中心）；RIKEN AIP（日本理化学研究所先进理工研究所）；Case Western Reserve University（凯斯西储大学）；The Hong Kong Polytechnic University（香港理工大学）；The University of Osaka（大阪大学）；University of Tokyo（东京大学）
Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference
Department of Computer Science, University of Manchester, UK（英国曼彻斯特大学计算机科学系）；Idiap Research Institute, Switzerland（瑞士Idiap研究所）；School of Computer Science, University of Sheffield, UK（英国谢菲尔德大学计算机科学学院）；National Biomarker Centre, CRUK-MI, University of Manchester, UK（英国曼彻斯特大学国家生物标志物中心）
Large Language Models Are Still Misled by Simple Bias Ensembles
Research Center for Social Computing and Interactive Robotics（社会计算与交互机器人研究中心）；Harbin Institute of Technology（哈尔滨工业大学）；Beijing Academy of Artificial Intelligence（北京人工智能研究院）
CARE-SD: Classifier-based analysis for recognizing and eliminating stigmatizing and doubt marker labels in electronic health records: model development and validation
Emory University（埃默里大学）；University of Florida（佛罗里达大学）；Children’s Healthcare of Atlanta（亚特兰大儿童医疗中心）；Mount Sinai（西奈山医院）；Georgia Institute of Technology（佐治亚理工学院）
