Summary of Distributional Preference Alignment Of Llms Via Optimal Transport, by Igor Melnyk et al.
Distributional Preference Alignment of LLMs via Optimal Transport
by Igor Melnyk, Youssef Mroueh, Brian Belgodere, Mattia Rigotti, Apoorva Nitsure, Mikhail Yurochkin, Kristjan Greenewald, Jiri Navratil, Jerret Ross
First submitted to arxiv on: 9 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: 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 paper proposes a novel method for aligning large language models (LLMs) on an unpaired preference dataset. The current alignment techniques rely on pairwise human preferences at the sample level, which does not imply distributional alignment. Alignment via Optimal Transport (AOT) addresses this limitation by making the reward distribution of positive samples stochastically dominant over the distribution of negative samples. AOT casts the first-order stochastic dominance as an optimal transport problem with a smooth and convex cost, allowing for a closed-form solution. The proposed approach is fine-tuned on LLMs, enabling alignment by penalizing the violation of stochastic dominance. Empirically, AOT achieves state-of-the-art results in the 7B family of models when evaluated with Open LLM Benchmarks and AlpacaEval. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps computers understand what we like or don’t like about language. Right now, these computers are not always on the same page as humans. The researchers propose a new way to make sure they’re aligned with our preferences. They call it Alignment via Optimal Transport (AOT). AOT makes sure that when computers see something good, they’ll prioritize it over something bad. This is important because computers need to understand what we like and don’t like to be useful. The researchers tested their approach on many different languages and found that it works well. |
Keywords
» Artificial intelligence » Alignment
