Scaled-up LLMs are more prone to sensible yet wrong answers

Researchers at the Universitat Politècnica de València, in Spain, have found that as Large Language Models (LLMs) grow larger and more sophisticated, they are more likely to span sensible yet wrong answers, making them less reliable. Researchers analyzed three popular LLMs; GPT by OpenAI, the LLaMA of Meta, and the BLOOM suite developed by BigScience, and noticed that with new versions accuracy increases, so do the hedging, refusal, or evasiveness. The study also found that LLMs rarely admit to a user that they do not know an answer.