Nvidia Paper Finds Irrelevant Text Weakens Models via Intrusion Heads

1 reports archived1 independent sourcesupdated 11/10/2026, 12:38:34

Synthesis & Latest Updates

1 Sources Cross-Validated

A new Nvidia-led paper finds that irrelevant text in RAG, tool-use and long-context prompts hurts models mainly by weakening their use of the question rather than directly suppressing the right answer. The authors identify a small set of 'intrusion heads' inside open models that read unrelated text and dilute the answer; masking about 10 such heads on 7B-scale models recovered 57% of the accuracy lost to distractors on average, versus a 97.6% oracle ceiling.

LATEST/A new Nvidia-led paper finds that irrelevant text in RAG, tool-use and long-context prompts hurts models mainly by weakening their use of the question rather than directly suppressing the right answer. The authors identify a small set of 'intrusion heads' inside open models that read unrelated text and dilute the answer; masking about 10 such heads on 7B-scale models recovered 57% of the accuracy lost to distractors on average, versus a 97.6% oracle ceiling.

Hourly heat curve

3 fully observed hours

Now

15.3

Peak

16.013:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

Rohan Paul (@rohanpaul_ai)T2Nov 10, 12:38·78 pts
Nvidia Paper Finds Irrelevant Text Weakens Models via Intrusion Heads
- Distractors hurt by weakening the model's use of the question, not by directly opposing the correct answer
- A small set of 'intrusion heads' inside models specifically reads irrelevant context
- Masking about 10 model-specific heads recovers 57% of lost accuracy on average with no training
View Related Reports & Analysis Original Source