Understanding what defines a good representation in large language models (LLMs) is fundamental to both theoretical understanding and practical applications. In this paper, we investigate the quality of intermediate representations in various LLM architectures, including Transformers and State Space Models (SSMs). We find that intermediate layers often yield more informative representations for downstream tasks than the final layers. To measure the representation quality, we adapt and apply a suite of metrics - such as prompt entropy, curvature, and augmentation-invariance - originally proposed in other contexts. Our empirical study reveals significant architectural differences, how representations evolve throughout training, and how factors like input randomness and prompt length affect each layer. Notably, we observe a bimodal pattern in the entropy of some intermediate layers and consider potential explanations tied to training data. Overall, our results illuminate the internal mechanics of LLMs and guide strategies for architectural optimization and training.
Exploring Mode Connectivity for Pre-trained Language Models
Tsinghua University(清华大学);Renmin University of China(中国人民大学);International Innovation Center of Tsinghua University(清华大学国际创新中心);Tencent Inc.(腾讯公司);Quan Cheng Laboratory(泉城实验室)
PhysicsArena: The First Multimodal Physics Reasoning Benchmark Exploring Variable, Process, and Solution Dimensions
The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));Beijing Future Brain Education Technology Co., Ltd.(北京未来脑教育科技有限公司);The Hong Kong University of Science and Technology(香港科技大学)
How Does Sequence Modeling Architecture Influence Base Capabilities of Pre-trained Language Models? Exploring Key Architecture Design Principles to Avoid Base Capabilities Degradation
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);iFLYTEK Co., Ltd(iFLYTEK公司)
Exploring Lightweight Large Language Models for Court View Generation
School of Computer Science and Engineering, Sun Yat-sen University(中山大学计算机科学与工程学院);School of Computer Science, South China Normal University(华南师范大学计算机学院);China Mobile Internet Co., Ltd.(中国移动互联网有限公司)
What does ChatGPT return about human values? Exploring value bias in ChatGPT using a descriptive value theory
Institute D’Or for Research and Teaching(多尔研究与教学学院);School of Psychology(心理学院);School of Information Management(信息管理学院);Dublin City University(都柏林城市大学);University of Western Australia(西澳大学)
本研究基于 Schwartz 基础价值观理论,通过 OpenAI API 重复提示 ChatGPT 生成文本并使用词袋方法分析,发现其输出无明显价值偏见且高保真传递了价值内容,仅存在社会导向价值观的融合。
Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation
Nanjing University(南京大学);Southeast University(东南大学);Shandong Women’s University(山东女子学院);State Key Laboratory for Novel Software Technology(现代软件技术国家重点实验室);National Institute of Healthcare Data Science(国家医疗数据科学研究院)
Dept. Computing Science & Alberta Machine Intelligence Institute (Amii), University of Alberta(计算科学系及阿尔伯塔人工智能研究所(Amii),阿尔伯塔大学);Canada CIFAR AI Chair, Amii(加拿大CIFAR人工智能主席,Amii)