Research map / Large language models / Retrieval and knowledge

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

NeurIPS 2025 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ParamMute, which suppresses mid-to-deep FFN layers whose activation correlates with unfaithful RAG outputs and then applies LoRA-based knowledge-augmented training plus preference optimization, evaluated on the new CoFaithfulQA benchmark and ConFiQA.

Problem

LLMs in RAG often produce outputs that contradict accurate retrieved context because internal parametric knowledge dominates generation; existing faithfulness methods focus on boosting use of external context and overlook internal knowledge.

Approach

Identify unfaithfulness-associated FFN layers (UA-FFNs) via activation-ratio gaps between faithful and unfaithful responses, suppress their activations (lambda=0), and then calibrate the model toward retrieved context with knowledge-augmented fine-tuning and a max-margin knowledge preference objective.

Key results

• On CoFaithfulQA, ParamMute achieves the best average MR among compared methods on most subsets, e.g., HotpotQA ConR 71.06 and MemR 6.17 versus SFT 70.92/6.24.
• On ConFiQA Question Answering subset, ParamMute reduces MemR to 3.69 (vs. Vanilla-RAG 38.51) with ConR 81.20.
• Suppressing 8 UA-FFN layers with lambda=0 plus adaptation yields the reported averages; the paper claims 6.17% and 54.63% faithfulness improvements on the two benchmarks.

Measured effects reported in the paper

Surprising findings

• Suppressing the UA-FFNs also lowers NLL on the faithful subset, suggesting it may not harm and could even help faithful generation.
• Higher self-consistency of parametric answers correlates with higher MemR and lower ConR, meaning confident parametric beliefs drive unfaithfulness.

Limitations

• Suppressing too many FFN layers (N around 10 or more) causes a sharp drop in context recall, indicating interference with other functions.
• Evaluation relies on closed-book self-generated parametric answers and LLM judges (GPT-4o and GLM-4-plus) for conflict labels, with ~90% human agreement.
• Experiments are mainly on LLaMA3-8B-Instruct with the controlled setting where documents are sufficient to answer questions.

Experimental setup

Datasets: Natural Questions SQuAD NewsQA TriviaQA SearchQA HotpotQA ConFiQA FaithEval

Benchmarks: CoFaithfulQA ConFiQA FaithEval

Metrics: ConR MemR MR NLL PPL

Baselines: Vanilla-RAG AttrPrompt O&I prompt COIECD SFT KAFT C-DPO DDR

Builds on: key-value memory view of FFNs DPO KAFT DDR Context-DPO

Closest papers by meaning

• To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts (ICLR 2025 · similarity 0.89)
• Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.89)
• FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows" (ICLR 2025 · similarity 0.88)
• Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness (ICLR 2025 · similarity 0.88)
• Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.88)
• Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.88)