人类将多样感知与语言输入转化为结构化行为的能力,被认为依赖于对概念的鲁棒表征学习。基于Transformer的大型语言模型(LLM)的快速发展,已涌现出多种与模型构建相关的计算要素,包括架构、微调方法、训练数据集等,但目前尚不清楚哪些要素对开发类人概念表征最为关键。此外,多数现有基准不适用于测量表征对齐,使得LLM在这些基准上的得分无法可靠评估其作为认知模型的进展。我们通过在 triplet 相似度任务上评估超过75个模型,解决了这些局限;该任务是认知科学中测量概念表征的成熟方法,使用来自 THINGS 数据库的概念。我们发现,指令微调与更大的注意力头维度是人类对齐的最强预测因素,而激活函数选择、多模态预训练及参数规模对对齐的影响有限。对齐得分与现有基准得分的相关性显示,虽然部分基准(如 BigBenchHard)比其他基准(如 MUSR)更能捕捉表征对齐,但没有一个基准能完全解释人与模型对齐的方差,证明了它们的不足。综上,我们的发现突出了推进LLM作为人类概念表征模型的关键计算要素,并填补了LLM评估中的一个关键空白。
英文摘要:
The human ability to translate diverse perceptual and linguistic inputs into structured behavior has been thought to rest on learning robust representations of concepts. The rapid advancement of transformer-based large language models (LLMs) has surfaced a diversity of computational ingredients relevant for model building - architectures, fine-tuning methods, and training datasets among others - yet it remains unclear which are most crucial for developing human-like conceptual representations. Further, most current benchmarks are ill-suited to measuring representational alignment, making LLMs' scores on them unreliable for assessing whether they are progressing as cognitive models. We address these limitations by evaluating over 75 models on a triplet similarity task, a method well established in cognitive science for measuring conceptual representations, using concepts from the THINGS database. We find that instruction fine-tuning and larger attention head dimensionality are among the strongest predictors of human alignment, while activation function choice, multimodal pretraining, and parameter size have limited influence on alignment. Correlations between alignment scores and existing benchmark scores reveal that while some benchmarks (e.g., BigBenchHard) better capture representational alignment than others (e.g., MUSR), none fully accounts for the variance in human-model alignment, demonstrating their insufficiency. Taken together, our findings highlight key computational ingredients for advancing LLMs as models of human conceptual representation and address a key gap in LLM evaluation.
LLM Factoscope: Uncovering LLMs' Factual Discernment through Inner States Analysis
Institute of Information Engineering, Chinese Academy of Sciences(中国科学院信息工程研究所);School of Cyber Security, University of Chinese Academy of Sciences(中国科学院大学网络空间安全学院)
MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs
School of Computer Science and Technology, East China Normal University(东华大学计算机科学与技术学院);Shanghai AI Lab, Shanghai Innovation Institute(上海人工智能实验室,上海创新研究院)
Uncovering Scaling Laws for Large Language Models via Inverse Problems
Singapore-MIT Alliance for Research and Technology(新加坡-麻省理工联盟研究技术中心);Dept. of Computer Science, National University of Singapore(新加坡国立大学计算机科学系);Agency for Science, Technology and Research(科技研究局);Institute of Data Science, National University of Singapore(新加坡国立大学数据科学研究院);SAP(SAP公司);AI Singapore(AI新加坡)
Uncovering inequalities in new knowledge learning by large language models across different languages
Peking University Shenzhen Graduate School(北京大学深圳研究生院);University of Science and Technology of China(中国科学技术大学);Wuhan University(武汉大学);Hong Kong University of Science and Technology(香港科技大学);Princeton University(普林斯顿大学);Microsoft Research(微软研究院);Microsoft Research Asia(微软亚洲研究院);Guanghua School of Management, Peking University(北京大学光华管理学院)
Analysing Differences in Persuasive Language in LLM-Generated Text: Uncovering Stereotypical Gender Patterns
Department of Computer Science, Aarhus University(阿arhus大学计算机科学系);AMD Silo AI;University of Tokyo(东京大学);IT University of Copenhagen(哥本哈根IT大学);Department of Computer Science, University of Copenhagen(哥本哈根大学计算机科学系)
Digital Skin, Digital Bias: Uncovering Tone-Based Biases in LLMs and Emoji Embeddings
University of North Texas(北德克萨斯大学);International Center for Artificial Intelligence Research and Ethics(国际人工智能研究与伦理中心);Southern Methodist University(南卫理公会大学);The University of Arizona(亚利桑那大学);The University of Alabama(阿拉巴马大学)
Identifying the Achilles' Heel: An Iterative Method for Dynamically Uncovering Factual Errors in Large Language Models
Renmin University of China(中国人民大学);The Chinese University of Hong Kong(香港中文大学);The Chinese University of Hong Kong, Shenzhen(香港中文大学深圳校区);Johns Hopkins University(约翰霍普金斯大学);Nanyang Technological University(南洋理工大学);Xiaohongshu Inc.(小红书公司);Tencent Inc.(腾讯公司)
When Ads Become Profiles: Uncovering the Invisible Risk of Web Advertising at Scale with LLMs
The University of New South Wales(新南威尔士大学);Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));Queensland University of Technology(昆士兰理工大学)
Uncovering Logit Suppression Vulnerabilities in LLM Safety Alignment
Huazhong University of Science and Technology(华中科技大学);Nanyang Technological University(南洋理工大学);University of New South Wales(新南威尔士大学);Nankai University(南开大学)
Uncovering Uncontrolled Repetition through Residual Stream Dynamics
Beijing University of Posts and Telecommunications(北京邮电大学);JIUTIAN Research(中移九天);Nanyang Technological University(南洋理工大学);Chongqing University of Posts and Telecommunications(重庆邮电大学)
Peering Inside the Black Box: Uncovering LLM Errors in Optimization Modelling through Component-Level Evaluation
Computer Science Department, KFUPM, Dhahran 31261, Saudi Arabia;SDAIA-KFUPM Joint Research Center for Artificial Intelligence, King Fahd University of Petroleum \& Minerals, Dhahran, 31261, Saudi Arabia
Tiny Brains, Giant Impact: Uncovering the Keystone Neurons of LLM with Just a Few Prompts
National University of Singapore(新加坡国立大学);Beijing University of Posts and Telecommunications(北京邮电大学);University of Science and Technology of China(中国科学技术大学)
Uncovering and Aligning Anomalous Attention Heads to Defend Against NLP Backdoor Attacks
Institute of Information Engineering, Chinese Academy of Sciences(中国科学院信息工程研究所);State Key Laboratory of Cyberspace Security Defense(网络空间安全防御国家重点实验室);School of Cyber Security, University of Chinese Academy of Sciences(中国科学院大学网络安全学院)
How well can a large language model explain business processes as perceived by users?
Eindhoven University of Technology(埃因霍温理工大学);IBM Research(IBM研究院);Peoples' Friendship University of Russia (RUDN University)(俄罗斯人民友谊大学(RUDN大学));Joint Institute for Nuclear Research(联合核子研究所);Vrije Universiteit Amsterdam(阿姆斯特丹自由大学);University of Skövde(舍夫德大学)
Layer by Layer: Uncovering Hidden Representations in Language Models
University of Kentucky(肯塔基大学);Mila-Quebec AI Institute(魁北克AI研究所);University of Montreal(蒙特利尔大学);New York University(纽约大学);University of California, Los Angeles(加州大学洛杉矶分校);Meta FAIR
From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization
School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences(中国科学院大学先进交叉学科学院);The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所认知与决策智能复杂系统重点实验室);School of Artificial Intelligence, University of Chinese Academy of Sciences(中国科学院大学人工智能学院);College of Computer Science, Inner Mongolia University(内蒙古大学计算机学院)