The University of Science and Technology of China (USTC) and Tencent YouTu Lab have released a research paper on a new framework called Woodpecker, designed to correct hallucinations in multimodal large language AI models. “Hallucination is a big shadow hanging over the rapidly evolving MLLMs,” writes the group, describing the phenomenon as when MLLMs “output descriptions that are inconsistent with the input image.” Solutions to date focus mainly on “instruction-tuning,” a form of retraining that is data and computation intensive. Woodpecker takes a training-free approach that purports to correct hallucinations from the basis of the generated text.

“The framework performs correction after a thorough diagnosis, incorporating a total of five stages: key concept extraction, question formulation, visual knowledge validation, visual claim generation, and hallucination correction,” VentureBeat writes, calling it a “groundbreaking approach.”

“Like a woodpecker heals trees, it picks out and corrects hallucinations from the generated text,” the group writes in their research paper, “Woodpecker: Hallucination Correction for Multimodal Large Language Models.” The researchers have released Woodpecker’s source code on GitHub, so others may build on the framework.

“The stages of Woodpecker work in harmony to validate and correct any inconsistencies between image content and generated text,” VentureBeat explains, noting that “first, it identifies the main objects mentioned in the text, then, it asks questions around the extracted objects, such as their number and attributes.”

The answers are provided using “expert models in a process called visual knowledge validation,” per VentureBeat.

Woodpecker then “converts the question-answer pairs into a visual knowledge base consisting of object-level and attribute-level claims about the image” and, finally, “modifies the hallucinations and adds the corresponding evidence under the guidance of the visual knowledge base,” VentureBeat adds.

Digital Information World calls Woodpecker “an absolute gamechanger for AI systems,” and describes the hallucination problem as “a huge hurdle linked to AI systems that hangs over MLLMs.”

Wired explores whether “chatbot hallucinations are poisoning web search.” Tidio, a company that trains customer service chatbots, produced a study that found about 86 percent of Internet users have experienced AI hallucinations, and 93 percent of users believe AI hallucinations can “cause harm” related to things like election manipulation, privacy risks and misinformation.

Based on a flurry of defamation lawsuits, Ars Technica suggests ChatGPT’s hallucinations may “ruin your life,” detailing the mishap of Australian regional mayor Brian Hood, who “made headlines by becoming the first person to accuse ChatGPT’s maker, OpenAI, of defamation.” Hood ultimately settled out of court.

Woodpecker “vows to make a huge difference in terms of enhancing accuracy and retaining accuracy along the way for AI systems,” Digital Information World explains.