New York University

January 14, 2011 - A study from New York University has revealed the potential risks of large-scale language models (LLMs) in medical information training. The study shows that even training data containing as little as 0.001% of incorrect information can cause the model to output inaccurate medical answers. Data "poisoning" is a relatively simple concept: LLM is usually trained on large amounts of text, mostly from the Internet. By injecting specific information into the training data, it is possible for the model to treat this information as fact when generating answers. This approach doesn't even require direct access to the LLM itself, just the purpose...