A recent study has found that while the latest large language models (LLMs) are becoming more accurate in solving complex problems, they are also becoming less reliable when it comes to admitting they don’t know an answer. This new research, conducted by artificial intelligence (AI) experts from the Universitat Politècnica de València in Spain, tested major AI models, including BigScience’s BLOOM, Meta’s Llama, and OpenAI’s GPT, revealing surprising insights into their capabilities and shortcomings.
In the study, researchers posed thousands of questions on subjects like maths, science, and geography to each of the models, evaluating their responses as correct, incorrect, or avoidant. The findings, published in the prestigious journal “Nature”, showed that while newer models improved in handling more challenging questions, they were also less likely to admit uncertainty when they couldn’t answer correctly. This marks a shift from earlier models, which would often indicate when they couldn’t find an answer or needed more information to provide a response.
The researchers noted that this tendency to guess rather than give an avoidant answer has raised concerns about the models’ overall reliability, particularly when dealing with simpler queries. While LLMs are designed to predict and generate content from large data sets, the study highlighted a troubling trend: despite their advanced capabilities, newer models sometimes make basic mistakes.
“Full reliability is not even achieved at very low difficulty levels,” the study revealed. “Although the models can solve highly challenging instances, they still fail at very simple ones.”
A notable example is OpenAI’s GPT-4, which demonstrated a sharp reduction in avoidant responses compared to its predecessor, GPT-3.5. However, the researchers were surprised to find that this did not correspond to an increase in the model’s ability to avoid answering outside its range of knowledge.
Despite the significant technological advancements in AI, the researchers concluded that there has been “no apparent improvement” in the newer models’ ability to solve basic problems.
The study’s findings raise questions about the future of AI and the development of more transparent, reliable language models. As AI technology continues to evolve, the need for systems that can recognize their own limitations is becoming increasingly important.