New Study Reveals How Irrelevant Context Degrades LLM Accuracy and How to Fix It
A new arXiv paper identifies a mechanism called the Intrusion Pathway that explains why language models lose accuracy when prompts contain irrelevant, off-topic context. The authors localize the problem to a sparse set o...