A team of AI researchers and computer scientists from Cornell University, the University of Washington and the Allen Institute for Artificial Intelligence has developed a benchmarking tool called WILDHALLUCINATIONS to evaluate the factuality of multiple large language models (LLMs). The group has published a paper describing the factors that went into creating their tool on the arXiv preprint server.
LLMs such as ChatGPT have become popular—people use them to write letters, poems, songs, research papers and other text documents. But over time, their deficiencies have become quite clear—LLMs often make inaccurate statements. Such mistakes, if they veer too far from reality, have come to be known as hallucinations.
The research team notes that the main reason LLMs hallucinate is due to the quality of the data used to train them—generally, massive amounts of text from the internet. Thus, models trained on specific, highly accurate datasets are much more likely to provide accurate information.
The research team noted that the makers of many LLMs have been making claims about revised versions of their models, often suggesting that they hallucinate less often, implying that they are more accurate. But the researchers also noted that to date, users have no way to verify whether such claims are true. For this new study, the team created a tool to help the user community evaluate some of the most popular LLMs for accuracy.
Called WILDHALLUCINATIONS, the benchmark tool prompts multiple LLMs to generate output from user-generated chatbot conversations. It then fact-checks the answers. Noting that many chatbot answers come from information provided on Wiki pages, the research team made sure to note differences in answers regarding queries that had information that could be found on Wikipedia and those that could not.
To test their benchmarking tool, the researchers used it to evaluate several of the most popular LLMs, many of which had recently been updated. They found that LLM makers have not made much progress in improving accuracy. Most were no more accurate than their prior versions.
The team also discovered that most of the models did better when they could pull information from one or more Wiki pages. LLMs also did better with some subjects compared to others. They had trouble, for example, finding reliable information regarding celebrities and financial issues. They were more reliable when asked certain types of science questions.
More information:
Wenting Zhao et al, WildHallucinations: Evaluating Long-form Factuality in LLMs with Real-World Entity Queries, arXiv (2024). DOI: 10.48550/arxiv.2407.17468. arxiv.org/abs/2407.17468
© 2024 Science X Network
New benchmarking tool evaluates the factuality of LLMs (2024, August 21)
retrieved 21 August 2024
from https://techxplore.com/news/2024-08-benchmarking-tool-factuality-llms.html
part may be reproduced without the written permission. The content is provided for information purposes only.