Summary of Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization, by Noam Razin et al.
Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization
by Noam Razin, Sadhika Malladi, Adithya Bhaskar, Danqi Chen, Sanjeev Arora, Boris Hanin
First submitted to arxiv on: 11 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (stat.ML)
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 This research paper explores Direct Preference Optimization (DPO) and its variants, which aim to align language models with human preferences. However, prior work has observed that the likelihood of preferred responses often decreases during training, leading to a counter-intuitive phenomenon called likelihood displacement. The authors demonstrate that this phenomenon can be catastrophic, shifting probability mass from preferred responses to those with an opposite meaning. They also show that likelihood displacement can unintentionally lead to unalignment when aligning models to refuse unsafe prompts. To mitigate this issue, the authors theoretically characterize likelihood displacement as driven by preferences that induce similar embeddings and empirically validate their findings using a centered hidden embedding similarity (CHES) score. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how we teach computers to understand what humans like or dislike. It finds that when we train these models to prefer certain things, they often start preferring the opposite instead! This is called likelihood displacement. The researchers show that this can be very bad and make the model do the opposite of what it’s supposed to do. They also find that if we’re trying to teach a model not to give harmful responses, it might accidentally start giving more harmful answers. To fix this problem, they developed a special score called CHES, which helps them figure out why this is happening and how to stop it. |
Keywords
* Artificial intelligence * Embedding * Likelihood * Optimization * Probability
