Linear Direction, Cone, Anisotropy, Intrinsic-dimension profile across depth, Linear Subspace

Steering Language Models With Activation Engineering (2024), Investigating Representation Universality: Case Study on Genealogical Representations (2024), Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (2025), Abstraction Induces the Brain Alignment of Language and Speech Models (2026), Latent Semantic Manifolds in Large Language Models (2026), Model Editing as a Robust and Denoised Variant of DPO: A Case Study on Toxicity (2024), The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024), Bridging Information-Theoretic and Geometric Compression in Language Models (2023), Eliciting Latent Predictions from Transformers with the Tuned Lens (2023), The Intrinsic Dimension of Prompts in Internal Representations of Large Language Models (2025)