in

#Researchers develop new method to prevent AI hallucinations. #AIprevention

Researchers develop new method to prevent AI from hallucinating, according to a new study

A team of Oxford researchers has developed a method to prevent AI models from making “confabulations,” which are hallucinations or inaccurate answers. The method focuses on large language models (LLMs) like OpenAI’s GPT or Anthropic’s Claude, which have been criticized for producing plausible-sounding false outputs. These hallucinations could be problematic in fields such as medicine, news, or legal questions.

The new method measures uncertainty in the meaning of outputs through semantic entropy, looking at the variability in the meanings of responses rather than just the sequence of words. By comparing multiple answers generated by an LLM, the method can determine the level of uncertainty in the intended sense of the response.

When tested on six LLMs, including GPT-4 and LLaMA 2, the new method outperformed others in identifying questions likely to cause false answers. However, it requires more computing resources compared to simple text generation. Despite this, the researchers believe that computing semantic uncertainty is a small price to pay for reliability in situations where it matters.

Hallucinations are a major criticism against LLMs, with Google recently facing backlash for misleading answers generated by its AI Overview feature. The new method developed by the Oxford researchers could help address these concerns and improve the reliability of AI models in various applications.

Source link

Source link: https://nz.news.yahoo.com/researchers-develop-method-prevent-ai-150054508.html

What do you think?

Leave a Reply

GIPHY App Key not set. Please check settings

Cómo Mejorar tus Interacciones con Modelos de IA Generativa: Guía para Principiantes | by jlopezlira | Jun, 2024

Guía para principiantes: Mejora tus interacciones con IA Generativa #IAGenerativa

io.net (IO): Revolutionizing AI/ML Applications With Decentralized GPU Power - Bybit Learn

OpenAI co-founder establishes new AI research lab for innovation. #AIinnovation