Summary of Refusal in Llms Is An Affine Function, by Thomas Marshall et al.
Refusal in LLMs is an Affine Function
by Thomas Marshall, Adam Scherlis, Nora Belrose
First submitted to arxiv on: 13 Nov 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed affine concept editing (ACE) approach allows for direct steering of language models’ behavior by intervening in activations. By decomposing model activation vectors into affine terms, ACE combines subspace projection and activation addition to reliably control refusal responses across prompt types. The method is evaluated using LLM-based scoring on a collection of harmful and harmless prompts, demonstrating more precise control over model behavior than existing methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Language models can be controlled by directly intervening in their activations, which is important for steering their behavior. This paper proposes an approach called affine concept editing (ACE) that does this. ACE works by breaking down the model’s activation vectors into smaller parts and then manipulating these parts to change the model’s behavior. The researchers tested ACE on several different models and found that it was able to reliably control their refusal responses. |
Keywords
» Artificial intelligence » Prompt
