Abstract
Large language models (LLMs) often exhibit undesirable behaviours, such as generating untruthful or biased content. Editing their internal representations has been shown to be effective in mitigating such behaviours on top of the existing alignment methods. We propose a novel inference-time editing method, namely spectral editing of activations (SEA), to project the input representations into directions with maximal covariance with the positive demonstrations (e.g., truthful) while minimising covariance with the negative demonstrations (e.g., hallucinated). We also extend our method to non-linear editing using feature functions. We run extensive experiments on benchmarks concerning truthfulness and bias with six open-source LLMs of different sizes and model families. The results demonstrate the superiority of SEA in effectiveness, generalisation to similar tasks, as well as computation and data efficiency. We also show that SEA editing only has a limited negative impact on other model capabilities.
Original language | English |
---|---|
Title of host publication | Advances in Neural Information Processing Systems 37 (NeurIPS 2024) Main Conference Track |
Editors | Amir Globerson, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub Tomczak, Cheng Zhang |
Publisher | Curran Associates Inc |
Pages | 56958-56987 |
Number of pages | 30 |
ISBN (Electronic) | 9798331314385 |
Publication status | Published - 16 Dec 2024 |
Event | The Thirty-Eighth Annual Conference on Neural Information Processing Systems - Vancouver Convention Center, Vancouver, Canada Duration: 10 Dec 2024 → 15 Dec 2024 Conference number: 38 https://neurips.cc/Conferences/2024 |
Publication series
Name | Advances in Neural Information Processing Systems |
---|---|
Publisher | Curran Associates, Inc. |
Volume | 37 |
ISSN (Print) | 1049-5258 |
Conference
Conference | The Thirty-Eighth Annual Conference on Neural Information Processing Systems |
---|---|
Abbreviated title | NeurIPS 2024 |
Country/Territory | Canada |
City | Vancouver |
Period | 10/12/24 → 15/12/24 |
Internet address |