Summary of More Rlhf, More Trust? on the Impact Of Preference Alignment on Trustworthiness, by Aaron J. Li et al.
More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness
by Aaron J. Li, Satyapriya Krishna, Himabindu Lakkaraju
First submitted to arxiv on: 29 Apr 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 study investigates the trustworthiness of Large Language Models (LLMs) aligned with general-purpose preference data through Reinforcement Learning From Human Feedback (RLHF). The authors evaluate LLMs’ performance across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. Surprisingly, they find that RLHF on human preferences doesn’t always guarantee trustworthiness and can even lead to reverse effects. To better understand the influence of fine-tuning data on individual trustworthiness benchmarks, the authors propose adapting efficient influence function based data attribution methods. The results underscore the need for more nuanced approaches to model alignment from both data and framework perspectives. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine if computers could learn like humans do. They would get better and better at understanding language. But, what if these computers started saying things that weren’t true or were mean? That’s why it’s important to make sure they’re trustworthy. In this study, scientists looked at how well computers learned from human feedback. They found that just because a computer is good at understanding language doesn’t mean it’s trustworthy. Sometimes, the more it learns, the less trustworthy it becomes! To fix this problem, the scientists came up with a new way to understand what makes a computer’s learning trustworthy or not. |
Keywords
» Artificial intelligence » Alignment » Fine tuning » Reinforcement learning from human feedback » Rlhf
