Summary of Aligning Large Language Models From Self-reference Ai Feedback with One General Principle, by Rong Bao et al.
Aligning Large Language Models from Self-Reference AI Feedback with one General Principle
by Rong Bao, Rui Zheng, Shihan Dou, Xiao Wang, Enyu Zhou, Bo Wang, Qi Zhang, Liang Ding, Dacheng Tao
First submitted to arxiv on: 17 Jun 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 research paper proposes a novel AI feedback framework to scale supervisory signals for large language models (LLMs). The existing methods rely on powerful LLMs, carefully designed principles, and are susceptible to position bias. To address these challenges, the authors introduce a self-reference-based approach that enables 13B Llama2-Chat to provide high-quality feedback under simple principles like “best for humanity”. The framework involves three stages: generating criticism of other answers based on its own response as a reference, determining which answer better fits human preferences according to the criticism, and using semantic perplexity to calculate preference strength differences. Additionally, the authors employ a self-consistency method to reduce position bias impact. The proposed method achieves significant advantages in benchmark datasets through reinforcement learning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about helping AI systems give good feedback to each other. Right now, this feedback is often based on how well an AI system can understand what humans want. But this can be tricky because it’s hard for AI to know what humans really mean. To solve this problem, the researchers created a new way for AI systems to learn from each other. They call it “self-reference-based”. It works by having one AI system generate its own response and then criticize other responses based on that. This helps the AI system understand what’s good or bad feedback. The team also found ways to reduce errors in this process. The results show that their method is effective, even with large language models like Llama2-Chat. |
Keywords
» Artificial intelligence » Perplexity » Reinforcement learning
