Summary of Understanding Knowledge Drift in Llms Through Misinformation, by Alina Fastowski and Gjergji Kasneci
Understanding Knowledge Drift in LLMs through Misinformation
by Alina Fastowski, Gjergji Kasneci
First submitted to arxiv on: 11 Sep 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 Large Language Models (LLMs) have become integral to our digital ecosystem, but their reliability becomes critical when exposed to misinformation. This study analyzes the susceptibility of state-of-the-art LLMs to factual inaccuracies in a QnA scenario, which can lead to “knowledge drift,” significantly undermining model trustworthiness. We evaluate factuality and uncertainty using Entropy, Perplexity, and Token Probability metrics. Experiments reveal that an LLM’s uncertainty increases up to 56.6% when answering incorrectly due to false information exposure. Repeated exposure to the same false information can decrease uncertainty again (-52.8%), potentially manipulating model beliefs and introducing drift from original knowledge. These findings provide insights into LLMs’ robustness and vulnerability to adversarial inputs, paving the way for developing more reliable LLM applications across various domains. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study looks at how well big language models do when they’re given false information. These models are really important in our digital world, but if they get bad info, it can mess up their answers. The researchers tested how much these models’ answers changed when they were given fake news. They found that the models got really unsure (up to 56.6%) when they gave wrong answers due to false information. But if they kept getting the same fake news, the models started to get more confident again (-52.8%). This study helps us understand how well these models do in real-life situations and what we can do to make them better. |
Keywords
» Artificial intelligence » Perplexity » Probability » Token
