This research aims to unravel how large language models (LLMs) iteratively refine token predictions through internal processing. We utilized a logit lens technique to analyze the model's token predictions derived from intermediate representations. Specifically, we focused on (1) how LLMs access and utilize information from input contexts, and (2) how positioning of relevant information affects the model's token prediction refinement process. On a multi-document question answering task with varying input context lengths, we found that the depth of prediction refinement (defined as the number of intermediate layers an LLM uses to transition from an initial correct token prediction to its final, stable correct output), as a function of the position of relevant information, exhibits an approximately inverted U-shaped curve. We also found that the gap between these two layers, on average, diminishes when relevant information is positioned at the beginning or end of the input context. This suggested that the model requires more refinements when processing longer contexts with relevant information situated in the middle. Furthermore, our findings indicate that not all layers are equally essential for determining final correct outputs. Our analysis provides insights into how token predictions are distributed across different conditions, and establishes important connections to existing hypotheses and previous findings in AI safety research and development.
Unraveling Media Perspectives: A Comprehensive Methodology Combining Large Language Models, Topic Modeling, Sentiment Analysis, and Ontology Learning to Analyse Media Bias
Institute of IT Management and Digitization Research (IFID)(信息管理与数字化研究学院)
A Large Language Model Guided Topic Refinement Mechanism for Short Text Modeling
Nanjing University of Posts and Telecommunications(南京邮电大学);Jiangsu High Technology Research Key Laboratory for Wireless Sensor Networks(江苏省无线传感器网络高技术研究重点实验室)
Tags for DAGs: Graph Refinement with Meta-Informed Relations
Computer Science Department, Technical University of Darmstadt(达姆施塔特技术大学计算机科学系);Hessian Center for AI (hessian.AI)(海德堡人工智能中心);German Research Center for AI (DFKI)(德国人工智能研究中心);Centre for Cognitive Science, Technical University of Darmstadt(达姆施塔特技术大学认知科学中心);Department of Mathematics and Computer Science, Eindhoven University of Technology(埃因霍温理工大学数学与计算机科学系)
RAIR: Retrieval-Augmented Iterative Refinement for Chinese Spelling Correction
Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所);University of Chinese Academy of Sciences(中国科学院大学);Zhongke Fanyu Technology Co., Ltd(中科泛语科技有限公司)
Training Language Model to Critique for Better Refinement
China Telecom Research Institute(中国电信研究院);The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University(清华大学对话人工智能(CoAI)小组);University of Electronic Science and Technology of China(电子科技大学);The Knowledge Engineering Group (KEG), Tsinghua University(清华大学知识工程小组);Zhipu AI(智谱AI)
Identifying and Decomposing Compound Ingredients in Meal Plans Using Large Language Models
Faculty of Health Sciences, University of Maribor(马里博尔大学健康科学学院);Faculty of Electrical Engineering and Computer Science, University of Maribor(马里博尔大学电气工程与计算机科学学院);The Nu B.V.(The Nu有限公司);Pain Research & Intervention Center of Excellence (PRICE), University of Florida(佛罗里达大学疼痛研究与干预卓越中心(PRICE));Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida(佛罗里达大学医学院健康结果与生物医学信息学系);Usher Institute, University of Edinburgh(爱丁堡大学厄舍研究所)
QUARTZ : QA-based Unsupervised Abstractive Refinement for Task-oriented Dialogue Summarization
Université de Lorraine, CNRS, Inria, LORIA(法国洛林大学、法国国家科学研究中心、法国国家信息与自动化研究所、LORIA研究所);Université Sorbonne Paris Nord, CNRS, Laboratoire d’Informatique de Paris Nord, LIPN(法国巴黎北大学、法国国家科学研究中心、巴黎北信息学实验室、LIPN研究所);Inserm, Centre de Recherche des Cordeliers, Université Paris Cité, Sorbonne Université(法国国家医学研究院、cordeliers研究中心、巴黎城市大学、索邦大学);HeKA, Inria Paris(HeKA、法国国家信息与自动化研究所巴黎分所);Department of Medical Informatics, Assistance Publique Hôpitaux de Paris, Georges Pompidou European Hospital(医学信息学系、巴黎公共医院、乔治·蓬皮杜欧洲医院)
本文提出无监督框架 QUARTZ,利用 LLM 生成任务问答并据此筛选高质量对话摘要,再微调模型,在零样本设置下达到接近完全监督 SotA 的效果。
Extracting Dataset Mentions in Forced Displacement and FCV Documents: A Weakly Supervised Framework with LLM-Based Label Refinement
Development Data Group(发展数据局);Office of the World Bank Group Chief Statistician(世界银行集团首席统计学家办公室);The World Bank(世界银行);World Bank UNHCR Joint Data Center(世界银行-联合国难民署联合数据中心)
Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement
Massachusetts Institute of Technology(麻省理工学院);Allen Institute for Artificial Intelligence(艾伦人工智能研究所);University of Washington(华盛顿大学);University of Southern California(南加州大学)
Efficient argument classification with compact language models and ChatGPT-4 refinements
Institute of Electronics(电子学研究所);AGH University of Krakow(克拉科夫AGH大学);Center for Collective Intelligence(集体智能中心);Massachusetts Institute of Technology(麻省理工学院);Faculty of Humanities(人文学院)