NVIDIA paper: masking ~10 'intrusion heads' recovers 57% of distractor accuracy loss
rohanpaul_ai · x · 2026-11-10
A NVIDIA-led paper, "The Intrusion Pathway," shows irrelevant context hurts LLMs mainly by weakening use of the question, not by directly suppressing the right answer. Corruption flows through MLPs into a sparse set of attention heads dubbed intrusion heads, which read unrelated text and erode answer support. These heads transfer across tasks and are largely weight-disjoint from evidence-integrating heads. On 7B-scale models across 4 tasks, a training-free mask on 10 heads (1% of attention heads) recovers 57% of distractor-induced accuracy loss on average; an oracle fix using a matched clean run recovers 97.6%. A practical, deployable mitigation for RAG and long-context serving with noisy retrieved context.
Related event: NVIDIA Paper Identifies 'Intrusion Pathway' Behind Irrelevant Context Harm(2 posts)→
More from Infra
- Chrome 153 ships native tab recorder: 120fps 4K AV1 capture, no ffmpeg needed — Vjeux · 2026-11-10
- AWS EKS now offers per-Pod hardware isolation with Nitro Sandbox — tekbog · 2026-11-10
- AI capex needs to grow 5-10x before oversupply, argues Rob LeClerc — robleclerc · 2026-11-10
- simdjson disables 256-bit LASX on Loongson 3A6000: 128-bit LSX is 6.7% faster — lemire · 2026-11-10
- Vera launches gyms, benchmarks and agents for cloud infrastructure operations — tianyin_xu · 2026-11-10
- Memory pricing itself out of demand: PC BOM share hits 35-38%, shipments plunge — tengyanAI · 2026-11-10