原帖引用 · 原文
New Nvidia + other labs paper shows that irrelevant text hurts mainly by weakening the model's use of the question, not by directly pushing against the right answer.
Prompts in RAG, tool use, and long-context apps often carry text that has nothing to do with the question. Even a line like "Computers are useful" lowers a model's support for the right answer.
They traced this inside open models and found intrusion heads, a small set of attention heads that read the unrelated text and weaken the answer. Masking those heads' attention to the context leaves the heads that use helpful evidence untouched.
Across 4 tasks on 7B-scale models, masking about 10 heads recovered 57% of the accuracy lost to distractors on average. An oracle fix that needs a matching clean run recovered 97.6%.
If you serve open-weight models with noisy retrieved context, a training-free mask on a few model-specific heads can win back much of the lost accuracy.