Summary of Throne: An Object-based Hallucination Benchmark For the Free-form Generations Of Large Vision-language Models, by Prannay Kaul et al.
THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models
by Prannay Kaul, Zhizhong Li, Hao Yang, Yonatan Dukler, Ashwin Swaminathan, C. J. Taylor, Stefano Soatto
First submitted to arxiv on: 8 May 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); 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 Mitigating hallucinations in large vision-language models (LVLMs) is an ongoing challenge. Recent benchmarks focus on hallucinations in specific question formats, neglecting those that occur in open-ended free-form responses, dubbed “Type I hallucinations”. Conversely, these benchmarks often rely on external API calls to models that may change, which can impact the accuracy of results. The study observes a curious phenomenon: reducing Type II hallucinations does not necessarily reduce Type I hallucinations; instead, the two forms are often anti-correlated. To address this, the authors propose THRONE, an automatic framework for evaluating Type I hallucinations in LVLM free-form outputs. By analyzing public language models and computing informative metrics, the study reveals that improving existing metrics does not reduce Type I hallucinations, and established benchmarks for measuring them are incomplete. The researchers also provide a data augmentation method to reduce both Type I and Type II hallucinations as a strong baseline. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper talks about how large models that can understand pictures and words (vision-language models) sometimes make up things they didn’t see or hear. This is called “hallucination”. The researchers looked at what happens when these models are asked to describe things in their own words, rather than just answering simple questions. They found out that there’s a problem with these models making things up in these free-form responses, which they call “Type I hallucinations”. The authors created a new way to measure this kind of hallucination and used it to test many recent language models. They discovered that just because a model is good at answering simple questions doesn’t mean it won’t make things up when describing something freely. |
Keywords
» Artificial intelligence » Data augmentation » Hallucination
