Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2403.05171 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866916316998270976 |
|---|---|
| author | Zhang, Xiaoying Ton, Jean-Francois Shen, Wei Wang, Hongning Liu, Yang |
| author_facet | Zhang, Xiaoying Ton, Jean-Francois Shen, Wei Wang, Hongning Liu, Yang |
| contents | We introduce Adversarial Policy Optimization (AdvPO), a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback (RLHF) for Large Language Models (LLMs). Over-optimization occurs when a reward model serves as an imperfect proxy for human preference, and RL-driven policy optimization erroneously exploits reward inaccuracies. In this paper, we begin by introducing a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model, without the need for computationally expensive reward ensembles. AdvPO then addresses a distributionally robust optimization problem centred around the confidence interval of the reward model's predictions for policy improvement. Through comprehensive experiments on the Anthropic HH and TL;DR summarization datasets, we illustrate the efficacy of AdvPO in mitigating the overoptimization issue, consequently resulting in enhanced performance as evaluated through human-assisted evaluation. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2403_05171 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Overcoming Reward Overoptimization via Adversarial Policy Optimization with Lightweight Uncertainty Estimation Zhang, Xiaoying Ton, Jean-Francois Shen, Wei Wang, Hongning Liu, Yang Machine Learning Artificial Intelligence We introduce Adversarial Policy Optimization (AdvPO), a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback (RLHF) for Large Language Models (LLMs). Over-optimization occurs when a reward model serves as an imperfect proxy for human preference, and RL-driven policy optimization erroneously exploits reward inaccuracies. In this paper, we begin by introducing a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model, without the need for computationally expensive reward ensembles. AdvPO then addresses a distributionally robust optimization problem centred around the confidence interval of the reward model's predictions for policy improvement. Through comprehensive experiments on the Anthropic HH and TL;DR summarization datasets, we illustrate the efficacy of AdvPO in mitigating the overoptimization issue, consequently resulting in enhanced performance as evaluated through human-assisted evaluation. |
| title | Overcoming Reward Overoptimization via Adversarial Policy Optimization with Lightweight Uncertainty Estimation |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2403.05171 |