Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models
Youngji Roh, Hyunjin Cho, Jaehyung Kim
摘要
Large Language Models (LLMs) exhibit highly anisotropic internal representations, often characterized by massive activations, a phenomenon where a small subset of feature dimensions possesses magnitudes significantly larger than the rest. While prior works view these extreme dimensions primarily as artifacts to be managed, we propose a distinct perspective: these dimensions serve as intrinsic interpretable functional units arising from domain specialization. Specifically, we propose a simple magnitude-based criterion to identify Domain-Critical Dimensions in a trainingfree manner. Our analyses reveal that such dimensions behave as interpretable semantic detectors for symbolic/quantitative patterns or domain-specific terms. In addition, we introduce Critical Dimension Steering, which applies activation steering exclusively to the identified dimensions. Empirical results show that this approach outperforms conventional wholedimension steering in domain adaptation and jailbreaking scenarios. 1 The following are multiple choice questions (with answers) about high school biology. You must respond with a single alphabet character. Energy is harvested during cellular respiration in stages. Which of the following correctly states which phase of cellular respiration harvests the most energy and the correct explanation why? A. The most energy is released during the Krebs cycle because it is here that pyruvate is completely ... B. ... both FADH2 and NADH are produced. Each of those molecules will release 2 ATPs ... C. ... phosphorylation because in addition to the phosphorylation of ADP into ATP ... D. ... during oxidative phosphorylation because H2O is completely broken down into H + and O2. Answer: 334 (Top 1) 1068 (Top 4) 2106 (Top 7) Act. Token Act. Token Act. Token 2048.0 <bos> 65.0 _the 43.8 H 324.0 The 64.0 _the 43.0 _ATP 147.0 _biology 62.0 _is 42.2 _ATP 139.0 Energy 61.0 _to 41.5 _respiration 138.0 _Krebs 59.5 _to 41.2 _energy 137.0 _cellular 58.8 _the 41.2 _ATP 132.0 _school 57.5 _the 41.2 _phosphorylation 130.0 _harvested 55.8 _that 40.8 _ADP 130.0 _cellular 55.0 _The 38.5 _NAD 130.0 _respiration 54.8 _in 38.2 _phosphorylation The following are multiple choice questions (with answers) about high school mathematics. You must respond with a single alphabet character. Suppose 5 different integers are randomly chosen from between 20 and 69, inclusive. What is the probability that they each have a different tens digit? A. 14 B. 13 C. 100052969 D. 250052969 Answer: The following are multiple choice questions (with answers) about high school mathematics. You must respond with a single alphabet character. An equilateral triangle has sides of 12 inches. What is the approximate area of the triangle?
问 Lune
尽管问,回答会引用原文。
Lune 会读完全文和相关的 Paper,每个回答都附上所依据的原文段落。
它引用的 Paper19
- Measuring Massive Multitask Language UnderstandingDan Hendrycks, Collin Burns, Steven Basart, Andy Zou 等ICLR 2021 · 7,905 次引用
- Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 次引用
- Inference-Time Intervention: Eliciting Truthful Answers from a Language ModelKenneth Li, Oam Patel, Fernanda B. Viégas, Hanspeter Pfister 等NeurIPS 2023 · 1,549 次引用
- SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language ModelsGuangxuan Xiao, Ji Lin, Mickaël Seznec, Hao Wu 等ICML 2023 · 1,493 次引用
- Sparse Autoencoders Find Highly Interpretable Features in Language ModelsRobert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart 等ICLR 2024 · 1,072 次引用
相关 Paper
- AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak DefenderWeixiang Zhao, Jiahe Guo, Yulin Hu, Yang Deng 等EMNLP 2025
- Detecting What Queries Seek: Steering LLM Safety with FFN Output Activation MonitoringXiaohao Luo, Ying Wei, Rui ZhaoACL 2026
- Fine-Grained Activation Steering: Steering Less, Achieving MoreZijian Feng, Tianjiao Li, Zixiao Zhu, Hanzhang Zhou 等ICLR 2026 · 6 次引用
- Enhancing Automated Interpretability with Output-Centric Feature DescriptionsYoav Gur-Arieh, Roy Mayan, Chen Agassy, Atticus Geiger 等ACL 2025
- Jailbreaking via Activation Dilution: Neuron-Guided Adaptive Decomposition of Language ModelsJiahang Wang, Chenglong Fu, Xiaojiang Du, Bo LuoCCS 2026
