Summary of Towards a Systematic Evaluation Of Hallucinations in Large-vision Language Models, by Ashish Seth et al.
Towards a Systematic Evaluation of Hallucinations in Large-Vision Language Models
by Ashish Seth, Dinesh Manocha, Chirag Agarwal
First submitted to arxiv on: 29 Dec 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
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 Vision-Language Models (LVLMs) have shown impressive performance in complex multimodal tasks. However, these models still struggle with hallucinations when asked to recognize or infer diverse visual entities from images for complex vision-language tasks. To address this challenge, researchers propose HALLUCINOGEN, a novel visual question answering (VQA) benchmark that uses contextual reasoning prompts as hallucination attacks to evaluate the extent of hallucinations in state-of-the-art LVLMs. The benchmark categorizes visual entities into salient and latent types, with salient entities being prominently visible objects and latent entities requiring domain knowledge or contextual reasoning for accurate inference. Hallucination attacks are designed for both entity types to assess hallucinations while performing various vision-language tasks, such as locating or reasoning about specific entities within an image. The proposed benchmark is evaluated on eleven LVLMs, including open-source models like LLaMA-3.2 and commercial models like Gemini, with two hallucination mitigation strategies across multiple datasets. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about fixing a problem in big AI models that can recognize what’s happening in pictures. Right now, these models sometimes make mistakes by imagining things that aren’t really there. To help fix this, the researchers created a new way to test these models called HALLUCINOGEN. This benchmark helps us understand how well these models do when they have to figure out what’s going on in a picture and answer questions about it. The researchers used two kinds of tests: one for easy-to-see things (like a car) and another for harder-to-spot things that need more knowledge or thinking. They tested eleven AI models, including some that are available online and others that are used by companies. This helps us understand what these models can do well and where they need to improve. |
Keywords
» Artificial intelligence » Gemini » Hallucination » Inference » Llama » Question answering
