Summary of Extending Activation Steering to Broad Skills and Multiple Behaviours, by Teun Van Der Weij et al.
Extending Activation Steering to Broad Skills and Multiple Behaviours
by Teun van der Weij, Massimo Poesio, Nandi Schoots
First submitted to arxiv on: 9 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY)
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 paper investigates the effectiveness of activation steering techniques to reduce risks from large language models’ capabilities. The researchers compare the impact of reducing performance on broad skills versus narrower skills and find that steering broader skills is competitive. They also experiment with steering models to exhibit different behaviors, such as myopia and wealth-seeking. The results show that combining multiple steering vectors into one is unsuccessful, but injecting individual vectors at different places in a model simultaneously is promising. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models can do bad things, like write scary stories or give biased advice. To stop this from happening, we need to control how the models work. This paper tries out a new way to do this called activation steering. They want to see if it’s better to make the model good at lots of things or just one specific thing. They also try making the model more or less clever and interested in money. The results show that it’s not so good to mix all the controls together, but if we use them separately, it works a bit better. |
