Summary of Spectral Editing Of Activations For Large Language Model Alignment, by Yifu Qiu et al.
Spectral Editing of Activations for Large Language Model Alignment
by Yifu Qiu, Zheng Zhao, Yftah Ziser, Anna Korhonen, Edoardo M. Ponti, Shay B. Cohen
First submitted to arxiv on: 15 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); 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 Large language models (LLMs) are prone to generating untruthful or biased content. To combat this, editing internal representations has been shown to be effective when combined with existing alignment methods. Our novel inference-time editing method, spectral editing of activations (SEA), projects input representations onto directions with high covariance with truthful demonstrations and low covariance with hallucinated ones. We extend our approach to non-linear editing using feature functions. We test SEA on benchmarks for truthfulness and bias with six open-source LLMs of varying sizes and families. Our results show that SEA outperforms other methods in effectiveness, generalisation, computation efficiency, and data efficiency. Additionally, we find that SEA editing has limited negative impact on other model capabilities. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Some language models can generate false or biased information. To fix this problem, we propose a new way to edit the internal workings of these models when they’re being used. Our method, called spectral editing of activations (SEA), helps move the model’s output in the direction of truthful information while keeping it away from fake data. We tested our approach with six different language models and found that it works better than other methods. Our results also show that SEA doesn’t harm the models’ abilities too much. |
Keywords
» Artificial intelligence » Alignment » Inference
