Summary of Right Now, Wrong Then: Non-stationary Direct Preference Optimization Under Preference Drift, by Seongho Son et al.
Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
by Seongho Son, William Bankes, Sayak Ray Chowdhury, Brooks Paige, Ilija Bogunovic
First submitted to arxiv on: 26 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 introduces a new approach to reinforcement learning from human feedback (RLHF) for Large Language Models (LLMs), called Non-Stationary Direct Preference Optimisation (NS-DPO). It addresses the limitation of current algorithms not accounting for temporal preference drift by using a Dynamic Bradley-Terry model that models preferences via time-dependent reward functions. NS-DPO applies exponential weighting to focus learning on more time-relevant datapoints, and is shown to be effective in fine-tuning LLMs in scenarios with drifting preferences. Theoretical analysis provides upper bounds on the estimation error caused by non-stationary preferences. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary RLHF aligns LLMs with human preferences, but these can change over time due to external factors. Current algorithms don’t account for this “temporal preference drift”. The paper proposes NS-DPO, a new way to fine-tune LLMs that takes into account changing preferences. It uses a special model and weighting system to focus on more relevant data points. This makes LLMs stay robust when preferences change, while still performing well when they don’t. |
Keywords
* Artificial intelligence * Fine tuning * Reinforcement learning from human feedback * Rlhf
