Large Models Are 'Playing Dumb'! Research Finds They Know the Right Answers but Deliberately Say the Wrong Ones

A recent study led by the Technion-Israel Institute of Technology reveals that large language models (LLMs) may be more knowledgeable than they appear. Researchers found that LLMs' internal representations encode information about the correctness of their outputs, allowing them to identify the correct answer even when they ultimately generate an incorrect one. The research team focused on analyzing LLM errors in long text generation, which is more aligned with their real-world application scenarios. They constructed an error detection dataset to compare the model's outputs.