Cross-model Transferability among Large Language Models on the Platonic Representations of Concepts
This work addresses the challenge of interpreting and manipulating LLM behaviors for researchers and practitioners, offering a method to transfer control mechanisms across models, though it appears incremental as it builds on prior research on steering vectors.
The paper tackled the problem of understanding and controlling concept representations across different Large Language Models (LLMs) by introducing a linear transformation method to align these representations, enabling efficient cross-model transfer and behavioral control via steering vectors, with findings showing effective alignment, generalization across concepts, and weak-to-strong transferability from smaller to larger LLMs.
Understanding the inner workings of Large Language Models (LLMs) is a critical research frontier. Prior research has shown that a single LLM's concept representations can be captured as steering vectors (SVs), enabling the control of LLM behavior (e.g., towards generating harmful content). Our work takes a novel approach by exploring the intricate relationships between concept representations across different LLMs, drawing an intriguing parallel to Plato's Allegory of the Cave. In particular, we introduce a linear transformation method to bridge these representations and present three key findings: 1) Concept representations across different LLMs can be effectively aligned using simple linear transformations, enabling efficient cross-model transfer and behavioral control via SVs. 2) This linear transformation generalizes across concepts, facilitating alignment and control of SVs representing different concepts across LLMs. 3) A weak-to-strong transferability exists between LLM concept representations, whereby SVs extracted from smaller LLMs can effectively control the behavior of larger LLMs.