PaperlayerView exact revision on arXiv

Source-linked AI summary

SCoNE: Selective Context-aware Neuron Editing for Robust Retrieval-Augmented Generation

Chaewon Kim, Seo Yeon Park

arXiv:2609.00689v1cs.CL

TL;DR

Retrieval noise can distract LLMs when RAG contexts mix informative and irrelevant documents, while existing fixes may require pipeline additions or retraining. SCoNE selectively strengthens context-aware FFN neurons identified by attribution and cross-input variability, using limited mining data without fine-tuning or inference-time overhead. Across benchmarks and backbones, it matches or surpasses strong baselines, though the effects of mining-dataset composition and retrieval conditions remain unclear.

  • Problem

    RAG performance is sensitive to mixed-quality retrieved documents, which can distract LLMs and degrade knowledge-intensive answers.

  • Method

    SCoNE strengthens context-aware FFN neurons selected jointly by attribution and cross-input variability, using 100 HotpotQA mining samples without fine-tuning or inference-time overhead.

  • Results

    SCoNE consistently outperforms competitive baselines across benchmarks, ranking first on six of seven datasets with Llama-3-8B-Instruct and improving over RAG by 3.75% on average.

  • Takeaways & Limitations

    Variability-based neuron mining provides a practical criterion for improving RAG robustness to retrieval noise through lightweight model editing.

  • Takeaways & Limitations

    Neuron mining uses HotpotQA samples, and the effects of mining-dataset composition and cleaner retrieval conditions remain unclear.

Abstract

from arXiv · show

Retrieval-Augmented Generation (RAG) is highly sensitive to retrieval noise: when retrieved documents mix informative and irrelevant context, LLMs are easily distracted, leading to hallucinations. To overcome this, we propose SCoNE (Selective Context-aware Neuron Editing), a training-free model editing approach that improves retrieval noise robustness by selectively strengthening context-aware FFN neurons that are identified by both high attribution and high cross-input variability. SCoNE requires only a small number of mining samples, no fine-tuning, and no inference-time overhead. Across various knowledge-intensive question-answering benchmarks and two LLM backbones, SCoNE consistently outperforms competitive baseline methods. Our code is available at https://github.com/HYU-ARK-Lab/SCoNE.

1 Introduction

RAG can reduce hallucinations by grounding generation in retrieved evidence, but mixed-quality retrieval often distracts LLMs. SCoNE addresses this challenge through lightweight, selective neuron editing without retraining.

  • Mixed relevant, partially relevant, and irrelevant documents can distract LLMs and degrade RAG answers.
  • Existing robustness strategies use prompting, context refinement, or fine-tuning, each introducing trade-offs in components, latency, compute, forgetting, or data requirements.
  • Model editing modifies few parameters without fine-tuning, but conventional methods assume the target knowledge is known before editing.
  • RAG requires neurons that respond to informative contexts while remaining unaffected by noisy ones.
  • SCoNE selects context-aware neurons using both high attribution and high cross-input variability, then strengthens them without fine-tuning or inference-time overhead.It uses 100 mining samples from HotpotQA and consistently outperforms competitive baselines across benchmarks.

2 Method

SCoNE mines context-aware FFN neurons from mixed gold and distractor evidence using attribution and cross-input variability, then strengthens the selected neurons during inference.

  • Neuron Mining: The mining dataset contains queries, ground-truth answers, and context sets combining gold contexts with distractor contexts that do not entail the answer.The method characterizes how individual FFN neurons engage with mixed evidence during inference.
  • Attribution: Attribution compares each neuron's activation on the query alone with its activation on the query augmented by the full context set.Integrated Gradients interpolates between these activations, with the integral approximated by a 20-step Riemann sum.
  • Variability: Variability measures how the current attribution differs from a running average over preceding instances, capturing variation across neighboring inputs.The running average uses a window of W preceding instances under a fixed traversal order.
  • High-Attribution and Variability Neuron Selection: Context-aware neurons are selected by intersecting the top-50 positive-attribution and top-50 variability neurons for each sample, then ranking neurons by aggregate selection frequency.The intersection captures neurons with both strong contribution and input-dependent responses.
  • Neuron Enhancement: SCoNE amplifies selected neurons at inference time by scaling their corresponding FFN weights, with α controlling the enhancement strength.This enhancement is applied after the context-aware neuron set has been identified.

3 Experimental Setup

The experiments mine neurons from HotpotQA and evaluate retrieval-augmented generation across seven QA datasets using realistic retrieval and generator configurations.

  • Neuron Mining Dataset: Neuron mining uses the first 100 HotpotQA training instances, with M = 2 gold contexts and K = 8 distractor contexts.These settings define the context composition used for mining.
  • Evaluation Dataset: Evaluation covers NQ, ASQA, SCIQ, TriviaQA, HotpotQA, TruthfulQA, and PopQA using BERGEN development splits and KILT Wikipedia documents.The setup retrieves five documents per question with SPLADE-v3.
  • Relevant vs. Irrelevant Context Analysis: Table 2 compares accuracy separately for Relevant questions, where a retrieved document contains the gold answer, and Irrelevant questions, where none does.This split evaluates behavior under answer-containing versus answer-absent retrieved contexts.
  • Implementation Details: Experiments use Llama-3-8B-Instruct and Qwen-2.5-7B-Instruct as generators, SPLADE-v3 as retriever, five retrieved documents, and non-overlapping 100-word chunks.All experiments use the BERGEN RAG framework and are conducted on a single NVIDIA H200 GPU.

4 Results

SCoNE delivers strong retrieval-noise robustness across generators and benchmarks, while ablations show that cross-input variability complements attribution and that moderate settings are stable.

  • Main Results: SCoNE ranks first on six of seven datasets with Llama-3-8B-Instruct and improves over vanilla RAG by 3.75% on average.It outperforms RetRobust by 3.73% and remains within 0.72% of PA-RAG without additional training.
  • Main Results: SCoNE’s improvements disappear with randomly selected neurons, which remain on par with vanilla RAG.Accuracy also remains stable across different 100-example HotpotQA mining samples.
  • Relevant vs. Irrelevant Context Analysis: SCoNE consistently outperforms RAG and IRCAN on relevant contexts and is most robust under irrelevant contexts, outperforming all baselines on NQ and HotpotQA.The same trend extends across controlled-noise evaluation.
  • Comparison of Variability Measures: SCoNE’s running-residual variability measure consistently outperforms variance, standard deviation, and mean absolute deviation across three datasets.Variance and standard deviation produce identical results because they induce the same neuron ranking.
  • Ablation on Neuron Selection Criteria: Attr+Var improves over Attr-only by 1.83% on NQ, 3.00% on SCIQ, and 2.00% on HotpotQA, whereas Var-only is insufficient.The results identify variability as a complementary signal rather than a standalone selection criterion.
  • Hyperparameter Analysis: Performance improves monotonically with enhancement strength α, peaks at α = 7, and remains stable across small-to-moderate W, N, and k.Extreme settings such as W = 10 or N = 1000 degrade performance, while the default configuration is best or near-best.

5 Related Work

Prior work addresses retrieval noise through prompting, context refinement, fine-tuning, and model editing, each with distinct pipeline or adaptation constraints.

  • Retrieval Noise Robustness in RAG: Retrieval-noise robustness methods include prompt engineering, retrieved-context refinement, and fine-tuning of the generator.Refinement methods include reranking, compression, clue selection, evidence extraction, and attention-based compression.
  • Model Editing for Context Utilization: Model editing changes a small number of parameters without additional training, but conventional methods assume the target knowledge is known beforehand.That assumption is mismatched with open-ended, query-dependent retrieved contexts in RAG.
  • Model Editing for Context Utilization: SCoNE extends knowledge-agnostic neuron-level contextual control by combining attribution strength with cross-input variability.This criterion is designed to identify neurons selectively responsive to retrieved evidence.

6 Conclusion

SCoNE improves RAG robustness to retrieval noise by selectively enhancing context-aware neurons identified through attribution and cross-input variability. Benchmark results show that this variability-based mining criterion matches or surpasses strong baselines.

  • Conclusion: SCoNE selectively enhances context-aware neurons identified through attribution strength and cross-input variability to improve RAG robustness to retrieval noise.The framework is presented as a practical neuron-mining approach for noisy retrieval.
  • Conclusion: Across various benchmarks, SCoNE matches or surpasses strong baselines, supporting variability-based neuron mining as a practical criterion.

Limitations

SCoNE transfers across diverse benchmarks, but its neuron mining is based only on HotpotQA and contexts containing both gold and distractor content, leaving broader transfer conditions unresolved.

  • Limitations: Neuron mining uses HotpotQA samples only, so the influence of mining-dataset characteristics on selected neurons and downstream behavior remains unclear.More challenging QA datasets may produce different neuron distributions and transfer properties.
  • Limitations: SCoNE’s neuron selection under cleaner retrieval settings containing only gold supporting documents remains unexamined.The current setting relies on contexts containing both gold-content and distractor-content.
  • Limitations: Future work should investigate how mining-dataset composition and retrieval conditions affect neuron mining and retrieval-noise robustness.

A.1 Neuron Mining Dataset Setting

Neuron mining uses the first 100 HotpotQA training distractor examples, converting all sample documents into context-provided prompts for attribution analysis.

  • A.1 Neuron Mining Dataset Setting: The mining set consists of the first 100 samples from HotpotQA’s training distractor split.
  • A.1 Neuron Mining Dataset Setting: Each mining sample uses all documents directly as input context, without an additional retriever, to simulate multi-document RAG.
  • A.1 Neuron Mining Dataset Setting: The attribution analysis compares a context-provided prompt format with a corresponding prompt format without retrieved context.

A.2 Baseline Implementation Details

The appendix specifies baseline reproduction settings and the inference-time RAG prompt materials used for evaluation and comparison.

  • A.2 Baseline Implementation Details: PA-RAG and RetRobust are reproduced using released checkpoints or author-provided training recipes across Llama-3-8B-Instruct and Qwen-2.5-7B-Instruct.
  • A.2 Baseline Implementation Details: CAD uses α = 0.5, while IRCAN selects 20 attribution-ranked candidate neurons, applies α = 7, and retains k = 5 final neurons.
  • A.2 Baseline Implementation Details: The inference-time RAG prompt uses a Background section containing the top-5 documents retrieved by the retriever.
  • A.2 Baseline Implementation Details: LLM-based evaluation uses GPT-5-mini as judge on NQ, SCIQ, and HotpotQA with Llama-3-8B-Instruct.

A.3 Inference-Time RAG Prompt

Appendix experiments examine prompt evaluation, neuron-selection controls, hyperparameter sensitivity, controlled retrieval noise, activation behavior, and out-of-domain effects. Across these tests, SCoNE’s gains are strongest under noise while remaining broadly stable under practical settings.

  • A.4 LLM-based Evaluation: SCoNE achieves the highest average LLM-judge score and best performance on two of three datasets.
  • A.5 Random Neuron Selection: Randomly selected neurons perform comparably to vanilla RAG, whereas SCoNE achieves the best performance with the same k = 5 and α = 7 settings.
  • A.6 Sensitivity to Mining Samples: 52.69 ± 0.44 match accuracy remains stable across random sets of 100 HotpotQA mining examples.
  • A.7 Effect of Enhancement Strength and Number of Selected Neurons: α = 7 gives the strongest overall results, while k = 5 and k = 15 yield comparable performance.
  • A.8 Effect of Candidate Pool Size: Top-50 offers the best overall candidate-pool trade-off: it leads on all three Qwen datasets by 0.7-2.3 points, although Top-20 is 0.1-0.8 points higher on Llama.
  • A.9 Effect of Context Window Size: Performance remains stable for W = 1, 2, 3, 5 but consistently degrades at W = 10, despite highly similar selected neuron sets.
  • A.10 Effect of Neuron Mining Dataset Size: N = 100 and N = 500 perform comparably, whereas N = 1000 reduces performance despite more mining data, supporting relatively small mining sets.
  • A.12 Controlled Noise Experiments: With added distractors, vanilla RAG drops steadily while SCoNE degrades more slowly; its gain grows with noise and is largely invariant to gold-document position.
Loading 2609.00689v1…