AI models fail when they have to connect too many facts at once

New research finds that large language models consistently fail at complex reasoning tasks. These tasks require connecting many pieces of information simultaneously. This limitation holds even when models get more computing power or more examples. It points to a fundamental problem in how these models are built.

Everyone assumed that with enough data and compute, AI would eventually reason like a human. This paper shows current models hit a hard wall when they need to link more than a few facts together to solve a problem. This means today's AI models are fundamentally unreliable for tasks that need deep scientific or analytical reasoning. Think drug discovery or complex engineering.