Scholarlink

Representation engineering is a recent approach which has been used to improve the safety of language models. By extracting vector representations of concepts such as ‘truthfulness’ or ‘toxicity’, the response of a model can be classified or controlled with respect to the concept. Since the representation engineering literature focuses on linear representations, we present a method to collect activations which can be used to fit an arbitrary non-linear representation. Furthermore, we introduce SurfaceRep, a method which controls concepts via a radial basis function network (RBFN) representation embedded in a low-dimensional space, to investigate whether a more complex, non-linear representation is able to form more accurate local approximations of concepts than linear representations. We evaluate SurfaceRep on several benchmarks for Artificial Intelligence (AI) safety and analyse the strengths and weaknesses of non-linear methods. We find that although the method does not consistently outperform linear methods, it remains useful as an interpretability tool for faithfully visualising representations of concepts.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Concept Control for LLM Safety Using Radial Basis Function Representations

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mark Amos,
  • Yang Song,
  • Maurice Pagnucco

摘要

Representation engineering is a recent approach which has been used to improve the safety of language models. By extracting vector representations of concepts such as ‘truthfulness’ or ‘toxicity’, the response of a model can be classified or controlled with respect to the concept. Since the representation engineering literature focuses on linear representations, we present a method to collect activations which can be used to fit an arbitrary non-linear representation. Furthermore, we introduce SurfaceRep, a method which controls concepts via a radial basis function network (RBFN) representation embedded in a low-dimensional space, to investigate whether a more complex, non-linear representation is able to form more accurate local approximations of concepts than linear representations. We evaluate SurfaceRep on several benchmarks for Artificial Intelligence (AI) safety and analyse the strengths and weaknesses of non-linear methods. We find that although the method does not consistently outperform linear methods, it remains useful as an interpretability tool for faithfully visualising representations of concepts.