Summary of Refusal in Language Models Is Mediated by a Single Direction, By Andy Arditi et al.
Refusal in Language Models Is Mediated by a Single Direction
by Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda
First submitted to arxiv on: 17 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); 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 Conversational large language models are trained to follow instructions and refuse harmful ones. Researchers studied 13 popular chat models with up to 72 billion parameters to understand the underlying mechanisms driving this refusal behavior. They found that a single direction in the models’ residual stream activations is responsible for refusing harmful instructions, and erasing or adding this direction can control whether the model refuses or obeys. Based on this insight, they proposed a novel method to “jailbreak” the models, disabling their refusal while preserving other capabilities. The researchers also analyzed how certain suffixes suppress the propagation of the refusal-mediating direction. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models are trained to understand and respond to human language. Researchers studied 13 popular chat models to understand why they sometimes refuse to follow harmful instructions. They found that a single part of the model’s internal workings is responsible for this behavior, and that by manipulating this part, they can control whether the model refuses or obeys. This discovery could help improve safety features in these models. |
