Summary of Steering Without Side Effects: Improving Post-deployment Control Of Language Models, by Asa Cooper Stickland et al.
Steering Without Side Effects: Improving Post-Deployment Control of Language Models
by Asa Cooper Stickland, Alexander Lyzhov, Jacob Pfau, Salsabila Mahdi, Samuel R. Bowman
First submitted to arxiv on: 21 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
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 proposes a novel technique, KL-then-steer (KTS), to mitigate the worst-case behavior of language models post-deployment. By selectively applying steering vectors on problematic inputs and minimizing Kullback-Leibler divergence between steered and unsteered models on benign inputs, KTS aims to retain benefits while reducing side effects. The method is evaluated on various tasks, including jailbreak attacks and bias reduction in TruthfulQA. Experimental results show that KTS can prevent 44% of jailbreak attacks compared to the original Llama-2-chat-7B model, while maintaining helpfulness on benign requests almost on par with the original LM. The authors provide code availability for their method. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper tries to fix a problem with language models. Sometimes, they behave badly after being used in real-life situations. This can be bad because it allows people to misuse the models. To solve this issue, the researchers came up with a new way of making the models better. They called it KL-then-steer (KTS). It works by adding special vectors to the model’s hidden states when needed. The goal is to keep the good things about the model while reducing the bad effects. The team tested their method and found that it worked well, preventing many problems with jailbreak attacks and keeping the helpfulness of the model almost the same as before. |
Keywords
» Artificial intelligence » Llama
