Parametric Stubbornness: Mechanistically Isolating the Layer Shift and Sparsity Gradient of RAG Knowledge Conflicts in Llama-3
Published in Preprint, Zenodo, 2026
This work studies how RAG knowledge conflicts emerge in Llama-3 by applying two-phase activation patching across minimal-pair conflicts. The paper demonstrates that hallucination behavior is driven by contextual attention override rather than simple parametric suppression, and identifies a sparse semantic bottleneck in the conflict pathway.
Recommended citation: Rajabi, H. (2026). "Parametric Stubbornness: Mechanistically Isolating the Layer Shift and Sparsity Gradient of RAG Knowledge Conflicts in Llama-3." Preprint, Zenodo.
Download Paper