Researchers discover a shortcoming that makes LLMs less reliable

Large language models can learn to mistakenly link certain sentence patterns with specific topics — and may then repeat these patterns instead of reasoning.