Researchers in China developed a hallucination correction engine for AI models
A team of scientists from the University of Science and Technology of China and Tencent’s YouTu Lab have developed a tool to combat “hallucination” by artificial intelligence (AI) models.
Hallucination is the tendency for an AI model to generate outputs with a high level of confidence that don’t appear based on information present in its training data. This problem permeates large language model (LLM) research, and its effects can be seen in models such as OpenAI’s ChatGPT and Anthropic’s Claude.
https://www.youtube.com/embed/gz6v2eOSZTE
The USTC/Tencent team developed a tool called “Woodpecker” that they claim is capable of correcting hallucinations in multimodal large language models (MLLMs).
This subset of AI involves models such as GPT-4 (especially its visual variant, GPT-4V) and other systems that roll vision and/or other processing into the generative AI modality alongside text-based language modeling.
According to the team’s preprint research paper, Woodpecker uses three separate AI models, apart from the MLLM being corrected for hallucinations, to perform hallucination correction.
These include GPT-3.5 turbo, Grounding DINO and BLIP-2-FlanT5. Together, these models work as evaluators to identify hallucinations and instruct the model being corrected to regenerate its output in accordance with its data.
To correct hallucinations, the AI models powering Woodpecker use a five-stage process that involves “key concept extraction, question formulation, visual knowledge validation, visual claim generation, and hallucination correction.”
Related: Humans and AI often prefer sycophantic chatbot answers to the truth — Study
The researchers claim these techniques provide additional transparency and “a 30.66%/24.33% improvement in accuracy over the baseline MiniGPT-4/mPLUG-Owl.” They evaluated numerous “off the shelf” MLLMs using their method and concluded that Woodpecker could be “easily integrated into other MLLMs.”
Source: Read Full Article
