As the IA language models become more and more sophisticated, they play a crucial role in the generation of text in various fields. However, guaranteeing the accuracy of the information they produce remains a challenge. Disinformation, involuntary errors and biased content can spread quickly, which has an impact on decision -making, public discourse and user confidence.
Google's Deepmind research division has unveiled a powerful AI fact verification tool Designed specifically for models of large languages ​​(LLMS). The tool, named SAFE (semantic precision and fact assessment), aims to improve the reliability and reliability of the content generated by AI.
Safe operates on a multifaceted approach, taking advantage of IA advanced techniques to analyze and meticulously check the factual allegations. The granular analysis of the system breaks down the information extracted from long -form texts generated by LLM into distinct and autonomous units. Each of these units undergoes a rigorous verification, with security using Google search results to perform a complete correspondence of the facts. What distinguishes security is its incorporation of reasoning into several stages, including the generation of research queries and the subsequent analysis of research results to determine factual accuracy.
During in -depth tests, the research team used its safety to check around 16,000 facts contained in the results given by several LLM. They compared their results to the checks of human facts (crowdsourcés) and found that security corresponded to the conclusions of specialists 72% of the time. In particular, in cases where differences have occurred, human precision has outperformed in complete safety, reaching a remarkable precision rate of 76%.
The advantages of security extend beyond its exceptional precision. Its implementation is estimated at approximately 20 times more profitable than relying on the verifications of human facts, making it a financially viable solution to treat the large quantities of content generated by the LLM. In addition, Safe's scalability makes it well suited to meet the challenges posed by the exponential growth of information in the digital age.
Although security represents a significant step forward for the development of LLMs, challenges remain. Ensure that the tool remains up to date with the evolution of information and maintaining a balance between precision and efficiency are current tasks.
Deepmind has rendered the code and reference data game accessible to the public on github. Researchers, developers and organizations can take advantage of its capacity to improve the reliability of the content generated by AI.
Deepen the LLM world and explore effective solutions for word processing problems using important languages ​​of languages, LLAMA.CPP, and the orientation library in our recent article “Optimization of word processing with LLM. Overview of Lama.CPP and advice.“”
