Summary of From Distributional to Overton Pluralism: Investigating Large Language Model Alignment, by Thom Lake et al.
From Distributional to Overton Pluralism: Investigating Large Language Model Alignment
by Thom Lake, Eunsol Choi, Greg Durrett
First submitted to arxiv on: 25 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
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 investigates the alignment process of large language models (LLMs) and its effects on their output distributions. It re-examines previous findings that alignment reduces response diversity, concluding that this reduction is mainly due to quality control and information aggregation. Alignment suppresses irrelevant content and shifts the distribution towards longer responses covering diverse information in a single answer. The study also finds that aligned models do not surface unique information that cannot be recovered from base models without fine-tuning. In-context examples and semantic hints can elicit similar responses from base LLMs as those from alignment-tuned LLMs. These results support the Superficial Alignment Hypothesis, showing that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models (LLMs) are powerful tools that can generate human-like text. But did you know that their outputs can change after being “aligned”? This paper looks at what happens when LLMs are aligned and how it affects their ability to provide helpful responses. The researchers found that alignment doesn’t really reduce the diversity of responses, but instead helps get rid of useless information and focuses on providing longer answers that cover a lot of ground. They also discovered that the unique things that aligned models can do aren’t actually new – base LLMs can do them too if they’re given the right hints. |
Keywords
» Artificial intelligence » Alignment » Fine tuning
