[2403.03867] On the Origins of Linear Representations in Large Language Models

Abstract:Recent works have argued that high-level semantic concepts are encoded "linearly" in the representation space of large language models. In this work, we study the origins of such linear representations. To that end, we introduce a simple latent variable model to abstract and formalize the concept dynamics of the next token prediction. We use this formalism to show that the next token prediction objective (softmax with cross-entropy) and the implicit bias of gradient descent together promote the linear representation of concepts. Experiments show that linear representations emerge when learning from data matching the latent variable model, confirming that this simple structure already suffices to yield linear representations. We additionally confirm some predictions of the theory using the LLaMA-2 large language model, giving evidence that the simplified model yields generalizable insights.

View PDF HTML (experimental) Abstract:Recent works have argued that high-level semantic concepts are encoded "linearly" in the representation space of large language models. In this work, we study the origins of such linear representations. To that end, we introduce a simple latent variable model to abstract and formalize the concept dynamics of the next token prediction. We use this formalism to show that the next token prediction objective (softmax with cross-entropy) and the implicit bias of gradient descent together promote the linear representation of concepts. Experiments show that…

saved by

• Timothy Kostolansky

related reading

• The Linear Representation Hypothesis and the Geometry of Large Language Modelsarxiv.org
• Verbalizable Representations Form a Global Workspace in Language Modelstransformer-circuits.pub
• The Linear Representation Hypothesis and the Geometry of Large Language Modelsarxiv.org
• A History of Large Language Modelsgregorygundersen.com
• The Geometry of Categorical and Hierarchical Concepts in Large Language Modelsarxiv.org
• [2412.06769] Training Large Language Models to Reason in a Continuous Latent Spacearxiv.org
• Symmetry in language statistics shapes the geometry of model representationsarxiv.org
• [2602.15029] Symmetry in language statistics shapes the geometry of model representationsarxiv.org
• [2210.13382] Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Taskarxiv.org
• [2310.06824] The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetsarxiv.org
• [2511.08579] Training Language Models to Explain Their Own Computationsarxiv.org
• Large Language Diffusion Modelsarxiv.org