Summary of Unfamiliar Finetuning Examples Control How Language Models Hallucinate, by Katie Kang et al.
Unfamiliar Finetuning Examples Control How Language Models Hallucinate
by Katie Kang, Eric Wallace, Claire Tomlin, Aviral Kumar, Sergey Levine
First submitted to arxiv on: 8 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 paper investigates how large language models (LLMs) hallucinate when faced with unfamiliar queries. The researchers found that LLMs’ hallucinated predictions are influenced by unfamiliar examples in their fine-tuning data, which introduce concepts beyond the base model’s knowledge scope. By modifying how these unfamiliar examples are supervised, the authors suggest that they can influence the model’s responses to unfamiliar queries, such as saying “I don’t know.” The study empirically validates this observation through controlled experiments on TriviaQA and MMLU datasets, involving SFT, RL, and reward model fine-tuning. The authors also explore RL finetuning strategies for improving the factuality of long-form model generations, finding that strategically controlling hallucinations can minimize negative effects. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at how large language models behave when they don’t know the answer to a question. The team found that these models tend to make up answers based on examples they’ve seen before, but not necessarily accurate ones. They also discovered that by changing how these models are trained, they can make them say “I don’t know” more often instead of making things up. The study tested this idea using different types of training and found that it works. |
Keywords
* Artificial intelligence * Fine tuning * Supervised
