AI models ignore facts you give them. Now developers know when.

Researchers have figured out why large AI models sometimes ignore the information you give them and sometimes follow it perfectly. This means developers can now predict when an AI will stick to its training or accept new facts, depending on the situation.

Developers building with large AI models faced a mystery: sometimes the AI would follow instructions and use provided documents, and other times it would stubbornly stick to its original training. This paper provides a map to that mystery, showing that the AI's behavior depends on how certain its training data is and how the task is framed. Now, developers can design systems that are more reliable, knowing when an AI is likely to ignore new information.