ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation

Pengcheng Huang and ten co-authors submitted the paper "ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation" to arXiv on February 21, 2025. The paper, classified under computational linguistics, addresses improving faithfulness in retrieval-augmented generation by suppressing specific feed-forward network parameters. It has undergone four revisions, with the latest version dated July 9, 2026. ParamMute represents a targeted approach to reducing hallucinated or conflicting model outputs when external retrieved context is used during generation.