Paper deep dive
Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model
Jaemin Kim, Jae O Lee, Sumyeong Ahn, Seo Yeon Park
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/3/2026, 12:18:38 AM
Summary
Neuro-RIT is a novel framework for Retrieval-Augmented Language Models (RALMs) that enhances robustness against noisy or irrelevant retrieved contexts by shifting from coarse-grained layer adaptation to precision-driven neuron alignment. The method uses attribution-based neuron mining to identify and disentangle neurons responsible for processing relevant versus irrelevant contexts, followed by a two-stage instruction tuning strategy that suppresses irrelevant context-aware neurons and optimizes relevant evidence distillation.
Entities (5)
Relation Signals (3)
Neuro-RIT → improvesrobustnessof → RALM
confidence 100% · we propose Neuro-RIT... to enhance the robustness of RALM
Neuro-RIT → utilizes → Attribution-based neuron mining
confidence 95% · Our method explicitly disentangles neurons... using attribution-based neuron mining.
Neuro-RIT → evaluatedon → KILT-NQ
confidence 90% · We evaluate our method on a diverse set of open-domain and multi-hop question answering benchmarks as follows: KILT-NQ
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Retrieval-Augmented Language Models (RALMs) have demonstrated significant potential in knowledge-intensive tasks; however, they remain vulnerable to performance degradation when presented with irrelevant or noisy retrieved contexts. Existing approaches to enhance robustness typically operate via coarse-grained parameter updates at the layer or module level, often overlooking the inherent neuron-level sparsity of Large Language Models (LLMs). To address this limitation, we propose Neuro-RIT (Neuron-guided Robust Instruction Tuning), a novel framework that shifts the paradigm from dense adaptation to precision-driven neuron alignment. Our method explicitly disentangles neurons that are responsible for processing relevant versus irrelevant contexts using attribution-based neuron mining. Subsequently, we introduce a two-stage instruction tuning strategy that enforces a dual capability for noise robustness: achieving direct noise suppression by functionally deactivating neurons exclusive to irrelevant contexts, while simultaneously optimizing targeted layers for evidence distillation. Extensive experiments across diverse QA benchmarks demonstrate that Neuro-RIT consistently outperforms strong baselines and robustness-enhancing methods.
Tags
Links
- Source: https://arxiv.org/abs/2604.02194v1
- Canonical: https://arxiv.org/abs/2604.02194v1
Trouble viewing inline? Open PDF directly →
Full Text
72,941 characters extracted from source content.
Expand or collapse full text
Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model Jaemin Kim 1 Jae O Lee 1 Sumyeong Ahn † 2 Seo Yeon Park † 1 Abstract Retrieval-AugmentedLanguageModels (RALMs) have demonstrated significant potential in knowledge-intensive tasks; however, they remain vulnerable to performance degradation when presented with irrelevant or noisy retrieved contexts.Existing approaches to enhance robustness typically operate via coarse-grained parameter updates at the layer or module level, often overlooking the inherent neuron-level sparsity of Large Language Models (LLMs). To address this limitation, we propose Neuro-RIT (Neuron-guided Robust Instruction Tuning), a novel framework that shifts the paradigm from dense adaptation to precision-driven neuron align- ment. Our method explicitly disentangles neurons that are responsible for processing relevant versus irrelevant contexts using attribution-based neuron mining. Subsequently, we introduce a two-stage instruction tuning strategy that enforces a dual capability for noise robustness: achieving direct noise suppression by functionally deactivating neurons exclusive to irrelevant contexts, while simultaneously optimizing targeted layers for evidence distillation.Extensive experiments across diverse QA benchmarks demonstrate that Neuro-RIT consistently outperforms strong baselines and robustness-enhancing methods. 1. Introduction While Large Language Models (LLMs) have achieved re- markable success, they still face significant limitations, including their tendency to produce hallucinations in knowledge-intensive tasks (e.g., question-answering) (Wang & Yu, 2025; Huang et al., 2025). To mitigate these problems, Retrieval-Augmented Language Models (RALM) (Lewis 1 Hanyang University 2 KENTECH. Correspondence to: Sumyeong Ahn<sumyeongahn@kentech.ac.kr>, Seo Yeon Park<seoyeon- park@hanyang.ac.kr>. Preprint. April 3, 2026. et al., 2020) provide a promising solution. RALM operates in two main steps: (1) invoking a retriever to search for and extract relevant contexts to a given query from external sources, and (2) leveraging the retrieved contexts to prompt LLMs, thereby generating more accurate and grounded re- sponses (Guu et al., 2020; Izacard et al., 2023; Gao et al., 2023; Fan et al., 2024). Despite their ability to support grounded generation, RALMs face a significant challenge: query irrelevant (or even incorrectly retrieved contexts) con- taining noisy retrieved contexts can mislead the LLM and consequently degrade its performance (Wang et al., 2023; Cuconasu et al., 2024; Chen et al., 2024). This is because LLMs are not explicitly trained to be robust against noisy or irrelevant contexts, and thus they may be distracted by such misinformation (Jiang et al., 2024; Lin et al., 2025). While prior work has mainly focused on refining retrieved context (Glass et al., 2022; Vig et al., 2022; Xu et al., 2023; Wu et al., 2025b), a more direct and increasingly promising direction is to enhance the robustness of the generator LLM itself (Wei et al., 2024). Recent studies have sought to im- prove robustness through instruction-tuning (Lin et al., 2023; Wei et al., 2024; Wu et al., 2025a). This approach encour- ages LLMs to explicitly learn how to better utilize retrieved context by following task-specific instructions e.g., disre- garding query irrelevant retrieved information. For exam- ple, Lin et al. (2023) adopted instruction tuning to optimize the LLM’s ability to process external knowledge and em- pirically demonstrated that the LLM learns to effectively distinguish and utilize relevant information within the con- text while remaining robust against distracting noise. However, existing works typically perform parameter up- dates at the layer or module level, or low-rank adaptation. While effective, these approaches may not fully leverage the inherent neuron-level sparsity of LLMs, overlooking the observation that only a limited subset of neurons is active for any given task or input (Li et al., 2022; Xu et al., 2025). Consequently, these methods might inadvertently affect un- related parameters, which is suboptimal for enhancing the robustness of RALM. This suggests that neuron-aware in- struction tuning—specifically targeting neurons responsible for distinguishing relevant information from retrieval noise— could offer a more precise and robust adaptation. 1 arXiv:2604.02194v1 [cs.CL] 2 Apr 2026 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model To this end, we proposeNeuro-RIT(Neuron-guided RobustInstructionTuning) where we aim to specify neu- rons that significantly contribute to processing relevant and irrelevant contexts separately, and then perform parameter updates based on these identified neurons to improve the robustness of RALM. Our method aligns with the inherent neuron-level sparsity of LLMs.Neuro-RITconsists of two phases: 1) context-aware neuron mining, and 2) two- stage neuron-guided instruction tuning. Specifically, we first identify distinct subsets of neurons responsible for pro- cessing relevant versus irrelevant contexts using attribution- based neuron mining. Subsequently, we proceed with a two-stage tuning process guided by these identified neurons. The first stage focuses on denoising irrelevant context-aware neurons. In this phase, we functionally deactivate neurons that contribute to process irrelevant contexts by imposing a hard constraint (i.e., instruction-tuning them to emit an End- of-Text (EOT) token) to decouple them from the generation process. The second stage targets noise filtering enhance- ment, where we instruction-tune neuron-level parameters to enforce the extraction of relevant evidence. Consequently, Neuro-RITequips RALMs with the dual capability of effectively suppressing irrelevant contexts while actively leveraging relevant evidence. Our contribution is summa- rized as follows: •We proposeNeuro-RIT, a novel framework that marks a paradigm shift from coarse-grained layer adap- tation to precision-driven neuron alignment to enhance the robustness of RALM, where we explicitly disentan- gle the neural mechanisms responsible for processing relevant versus irrelevant contexts. •We introduce a two-stage neuron-guided instruction tuning strategy that enforces a dual capability for noise robustness, by achieving direct noise suppression via functionally deactivating irrelevant context-aware neu- rons while optimizing neuron-level parameters for evi- dence distillation. • We validateNeuro-RITthrough extensive experi- ments on both in-domain and out-of-domain datasets and demonstrate thatNeuro-RITconsistently out- performs strong baselines in knowledge-intensive QA tasks, proving its efficacy in mitigating the impact of retrieval noise without compromising generative capa- bility. 2. Related Work Refining Retrieval Contexts.Prior works have proposed various context refinement methods to enhance the quality of retrieved knowledge. For example, reranking retrieved documents (Glass et al., 2022), extracting query-related content from the retrieved documents (i.e., summarization) (Vig et al., 2022; Xu et al., 2023), and iterative retrieval along with intermediate reasoning paths (Trivedi et al., 2023; Wang et al., 2024; Wu et al., 2025b) have been explored. Re- cently, Wu et al. (2025b) proposed an advanced knowledge refinement method that incorporates re-ranking signals in generating Chain-of-Thought (CoT)-based summarization. However, many existing works have created extra modules for the knowledge refinement, which introduce additional steps beyond generation on RALM. These additional steps introduce extra computational costs, making it less practical for large-scale deployment (Wu et al., 2025a). Moreover, these additional steps may suffer from cascading errors (Yu et al., 2022), which ultimately limit their effectiveness in real-world applications. Robustness to Retrieved Contexts in RALM. Previ- ous studies have endeavored to improve the robustness of RALM to noisy or irrelevant retrieved contexts. Yoran et al. (2024) enhanced model robustness implicitly by training lan- guage models to directly generate correct answers despite input noise. Instead of relying solely on implicit learning, re- cent works utilize instruction-following capabilities through instruction tuning to explicitly articulate the denoising pro- cess. For instance, Wei et al. (2024) prompts the model to synthesize denoising rationales that explicitly analyze the retrieved contexts and articulate the reasoning path to the ground-truth answer, thereby allowing the model to learn how to filter out irrelevant or noisy contexts. Similarly, Wu et al. (2025a) enhances robustness against noisy con- texts by performing instruction tuning with high-quality data constructed via a citation rewrite mechanism, which utilizes NLI models to rigorously verify and refine the cita- tions, ensuring the model learns to cite only valid evidence. However, a limitation remains regarding the granularity of adaptation. By relying on dense optimization strategies such as full-parameter fine-tuning or layer-level parameter- efficient fine-tuning, existing approaches overlook the po- tential of neuron-level sparsity. Consequently, they neglect to selectively update the specific neurons critical for noise robustness. 3. Methodology ProblemFormulationandOverview. Retrieval- Augmented Language Models (RALMs) for knowledge- intensive tasks, such as question-answering, aim to generate answers using a generator LLM, based on a query and documents returned by a retriever. Formally, letD Q = (q i ,d i,j k j=1 ) N i=1 be a set of open-domain questions, whereq i denotes a query andd ij k j=1 is the set of top-kretrieved documents associated withq i . Note that the documents are ranked based on relevance score with the given query. Then, a generator LLMMis expected to produce the outputˆyas follows:ˆy i = M(Prompt i ) 2 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model ...... Query, Relevant, Irrelevant Retrieved Contexts :3 :2 :2 :1 :1 ퟎ.ퟗퟖ>ퟎ.ퟗퟏ>ퟎ.ퟖퟒ ퟎ.ퟖ>ퟎ.ퟕ> ퟎ.ퟕퟏ ퟎ.ퟖퟒ> ퟎ.ퟖퟎ> ퟎ.ퟕퟗ ... (풒 ퟏ ,풅 ퟏ 풊풓풆풍 ) (풒 ퟐ ,풅 ퟐ 풊풓풆풍 ) (풒 풏 ,풅 풏 풊풓풆풍 ) ... Select neurons with frequently occur across all data samples Extract neurons of highest attribution score for each sample Decoupling Context-Aware Neurons (풒 ퟏ ,풅 ퟏ 풓풆풍 ) ෩ 푷 풓풆풍 ෩ 푷 풊풓풆풍 ... ퟎ.ퟗퟐ>ퟎ.ퟖퟕ>ퟎ.ퟖퟓ Attr Score Attr Score Attr Score Attr Score :3 :3 1) Denoising Irrelevant Context-Aware Neurons ( ) Phase 1: Context-Aware Neuron MiningPhase 2: ...... : Top-3 layers (푷 풔풉풂풓풆풅 , 푷 풊풓풆풍 ) : Gradient Mask Relevant Context Aware Neurons Irrelevant Context Aware Neurons 푷 풓풆풍 푷 풊풓풆풍 푷 풔풉풂풓풆풅 : Relevant Contexts: Irrelevant Contexts 2) Active Noise Filtering Enhancement FFN Attention : TransformerLayer (FFN / Attention) Fine-tuning ONLY 푷 풊풓풆풍 Full FT is applied to the top-3 layers with the highest concentration of 푷 풔풉풂풓풆풅 , 푷 풊풓풆풍 neurons Layer 풏 : 0 16 17 18 29 30 31 ...... All neuron groups (푷 풓풆풍 , 푷 풔풉풂풓풆풅 , 푷 풊풓풆풍 ) are updated with group-specific masks. Full Fine-Tuning Neuron-Guided Instruction Tuning Figure 1. The overview of the Neuro-RIT framework. Phase 1: We identify neurons highly responsive to relevant and irrelevant contexts via attribution scores aggregated across samples, and decouple them into distinct sets: relevant (P rel ), irrelevant (P irrel ), and shared (P shared ) neurons. Phase 2: We proceed with a two-stage instruction tuning strategy. First, we instruction-tuneP irrel to encourage early emission of an End-of-Text (EOT) token, suppressing responses driven by irrelevant contexts. Second, we apply neuron-guided tuning with group-specific gradient masks, while performing full fine-tuning on the top-3 layers with the highest neuron density of P shared and P irrel . wherePrompt i = (q,d i,j k j=1 )refers to a predefined prompt template along with inputs. However, the retrieved documents setd i,j k j=1 often contain query-irrelevant documents that lead to performance degradation.To this end, we propose to identify relevant and irrelevant context-aware neurons and enhance their robustness to retrieved information. Our method consists of two primary components: 1) context-aware neuron mining, and 2) robustness enhancement instruction tuning on the identified neurons. 3.1. Context-Aware Neuron Mining Attribution-Based Context-Aware Neuron Identification. To quantify the contribution of each neuron to processing relevant and irrelevant contexts, we leverage the attribution method of Shi et al. (2024). For each queryq i , we first construct a triplet(q i ,d rel i ,d irrel i ). Here,d rel i denotes a top- ranked relevant document, andd irrel i denotes an irrelevant document with a zero relevance score. Next, for every query q i and documentd c i with context typec ∈ rel,irrel, we measure the attribution score of every neuronn(that exists in FFNs) to the predictiony. This is done via Integrated Gradients (Sundararajan et al., 2017). Specifically, letv(q i ) denote the activation of neuronngiven only the query, and v(q i ,d c i ) denote the activation given both the query and the document of context typec. We compute the attribution score as follows: A(n;q i ,d c i ) = v(q i ,d c i )−v(q i ) Z 1 α=0 ∂P(y | q i ,d c i ,v α ) ∂v α dα, (1) wherev α = v(q i ) + α v(q i ,d c i )− v(q i ) interpolates be- tween the two activation states forα∈ [0, 1], and the inte- gral is approximated using a 20-step Riemann sum. Based on these scores, for each query, we filter for the top 10% and select the top 20 highest-scoring neurons. We then aggregate these neuron selections across all queries within each context type, and identify the top-tmost frequently occurring neurons. We denote these initial candidate sets of neurons as ̃ P irrel and ̃ P rel for irrelevant and relevant contexts, respectively (See Figure 1). Decoupling Context-Aware Neurons. Once we obtain ̃ P rel and ̃ P irrel , we explore the intersection between the two sets. We observe that identified neuron sets ̃ P irrel and ̃ P rel re- veal a significant overlap 1 . This is natural since parameters associated to neurons in LLMs form distributed and overlap- ping representations rather than being functionally exclusive (Leng & Xiong, 2025). However, such overlapping neu- rons do not provide discriminative signals for identifying context-specific awareness, as their activations cannot be uniquely attributed to either relevant or irrelevant contexts. 1 Initial mining yields 7,380 relevant and 7,036 irrelevant context- aware candidate neurons with a significant overlap of 6,240 neu- rons. 3 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model To address this, we define distinct neuron sets by exclud- ing neurons that exhibit high attribution in both context types. Specifically, we define the overlapping neurons set asP shared = ̃ P rel ∩ ̃ P irrel , and remove them from each set as follows:P rel ← ̃ P rel shared ,P irrel ← ̃ P irrel shared . This procedure ensures that the resulting neuron sets capture dis- tinct context-specific awareness for relevant and irrelevant contexts, respectively (See Figure 1). Our mining results in 100 neurons forP rel andP irrel and 30 neurons forP shared . 3.2. Neuron-Guided Instruction Tuning Based on the identified neurons, we propose a two-stage neuron-guided instruction tuning framework: 1) denoising irrelevant context-aware neurons, and 2) noise filtering en- hancement for identified neurons. Denoising Irrelevant Context-Aware Neurons. Since P irrel are identified as primarily responsible for processing irrelevant or misleading contexts, allowing them to partici- pate in processing relevant information poses a risk of latent interference. Hence, we propose to instruction-tune these neurons to emit an End-of-Text (EOT) token, even when presented with a query and its relevant contexts accom- panied by an explicit instruction, only leveraging relevant information from the retrieved contexts. According to this, we impose a hard constraint that decouples these irrelevant context-aware neurons from generation. Noise Filtering Enhancement on Neuron-level Param- eters. We explore the distribution of irrelevant context- aware neurons that are included in bothP irrel andP shared across layers. Interestingly, we observe that these neurons tend to aggregate in specific layers (i.e., the last three lay- ers, 29, 30, and 31 for LLaMA-3-8B-Instruct), indicating that a subset of layers is primarily responsible for process- ing irrelevant contexts (See §4.5). Hence, we propose to instruction-tune these layers along with specified neurons groups,P rel ,P irrel , andP shared , to enforce robustness of RALM. Specifically, the instruction-tuning process is config- ured to teach the generator LLM, a particular interpretation: when presented with an original query and potentially noisy retrieved contexts, the LLM should treat this input as an instruction 2 to generate only the distilled, relevant summary. This effectively aligns the LLM’s generative capabilities with our context-filtering objective. For this, we construct a dataset as follows: for a pair(q i ,d i,j k j=1 )∈D Q , we em- ploy an external LLM to generate Relevant Summaries (RS) y RS i , which includes only query-relevant information from d ij k j=1 by giving a specific instruction 2 . In our experi- ments, we usedGPT-4.1.mini(OpenAI 2025) 3 though any LLM can be used. We construct an instruction-tuning 2 See Table A.3 in Appendix for details 3 https://openai.com/index/gpt-4-1/ dataset as follows: D RS = q i ,d ij k j=1 ), y RS i n i=1 We optimize neuron-level parameters by minimizing the next-token prediction loss on the output segment. We selectively update the parameters by applying gradient masks following Xu et al. (2025).Once done train- ing, to enable open-domain question answering, we de- sign a dual-instruction prompt consisting of (1) relevant summary extraction (inst RSE :extract relevant information from provided documents) and (2) a question answering instruction (inst QA :Answer a question as briefly as possible).Each training inputx i is formatted as a prompt combining these components: Prompt i = (inst RSE ,inst QA ,d ij k j=1 ,q i ) The instruction-tuned generator LLM produces the cor- responding answer asˆy i = M(Prompt i ). This design jointly enables the model to filter noisy documents, extract and align relevant evidence, and generate factually grounded answers. 4. Experiments 4.1. Data We evaluate our method on a diverse set of open-domain and multi-hop question answering benchmarks as follows: KILT-NQ (Natural Questions) (Kwiatkowski et al., 2019) is an open-domain question answering dataset constructed from real user queries and grounded in Wikipedia arti- cles. We use the dataset aligned with the KILT format, following (Rau et al., 2024). ASQA (Answer Sentence Question Answering) (Stelmakh et al., 2022) is a question answering benchmark that requires generating long-form, descriptive answers by synthesizing evidence from multi- ple sentences rather than extracting a single span. KILT- TriviaQA (Joshi et al., 2017) is a KILT-formatted version of TriviaQA, consisting of factoid questions paired with sup- porting Wikipedia documents. SCIQ (Welbl et al., 2017) is a domain-specific multiple-choice question answering dataset constructed from elementary-level science textbooks. POPQA (Mallen et al., 2023) is a benchmark composed of questions whose answers may change over time, designed to evaluate a model’s ability to update and rely on current infor- mation. KILT-HotpotQA (Yang et al., 2018) is a multi-hop question answering dataset that requires reasoning over two or more documents to derive the correct answer. Provided in the KILT format, each question is associated with multiple evidence documents. 2WikiMultiHopQA (Ho et al., 2020) is a multi-hop question answering dataset that explicitly requires connecting information across different Wikipedia articles. Each question involves multiple entities, making single-document reasoning insufficient. 4 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model 4.2. Baselines To validate the effectiveness ofNeuro-RIT, we conduct a comparative evaluation against three distinct categories of baselines: 1) Standard RALM, 2) Robustness Enhancing Tuning for RALMs, and 3) Retrieved Context Refinement. For the refinement methods, we additionally evaluate them in a compositional manner by attaching each refinement module to every baseline in the first two categories. Standard RALM.This baseline follows the conventional Retrieval Augmented Generation (RAG) paradigm, where retrieved documents are concatenated along with a given query without any additional control mechanisms. Robustness Enhancing Tuning for RALMs.These meth- ods improve robustness primarily by tuning the generator LLM itself when faced with irrelevant or misleading re- trieved contexts. • RetRobust 4 (Yoran et al., 2024) improves robustness implicitly by training the generator LLM to directly generate the correct answer even when the input con- tains distracting or misleading retrieved contexts. By exposing the LLM to noisy retrieval scenarios during training, it encourages the generator to be less sensitive to irrelevant evidence without explicitly supervising a denoising procedure. •InstructRAG 5 (Wei et al., 2024) improves robustness via instruction tuning with denoising behaviors. Con- cretely, the generator LLM is prompted to produce intermediate denoising rationales that analyze the re- trieved contexts, identify valid evidence, and then de- rive the final answer. This baseline represents a strong instruction-following approach that teaches the gen- erator how to ignore noise through explicit reasoning traces. • PA-RAG 6 (Wu et al., 2025a) enhances robustness through instruction tuning the generator LLM on high- quality data constructed by a citation rewrite mecha- nism. It leverages external verification (e.g., NLI-based checking) to refine citations and ensure that the supervi- sion emphasizes only valid supporting evidence. As a result, the LLM learns to ground responses on reliable contexts and reduce reliance on spurious or incorrect retrievals. 4 https://huggingface.co/Ori/llama-2-13b-peft-nq-retrobust 5 https://huggingface.co/meng-lab/2WikiMultiHopQA- InstructRAG-FT 6 https://huggingface.co/wuqiong1/PA-RAGMeta-Llama-3-8B- Instruct Retrieved Contexts Refinement.In contrast to generator- centric robustness tuning, these baselines aim to improve robustness by refining the retrieved contexts before they are consumed by the generator LLM. While effective, such methods typically introduce additional modules and extra inference steps (e.g., re-ranking or summarization), poten- tially increasing computational overhead and susceptibility to cascading errors. • Reranker 7 (Glass et al., 2022). This baseline employs a cross-encoder re-ranking model to reorder retrieved documents according to relevance to the query, and then selects top-ranked documents for prompting the generator. By filtering or prioritizing evidence at the retrieval stage, it reduces the chance that irrelevant contexts dominate the prompt. •RankCoT 8 (Wu et al., 2025b). RankCoT performs ad- vanced context refinement by incorporating re-ranking signals into CoT-based summarization of retrieved doc- uments. It generates a condensed and query-focused representation of the evidence through a reasoning- guided summarization process, and the generator then answers based on the refined summary. This baseline represents state-of-the-art refinement that can mitigate noise, but requires additional reasoning and processing beyond standard RAG. 4.3. Implementation Details All experiments are conducted using the generator LLM as LLaMA-3-8B-Instruct (Dubey et al., 2024), and training is performed on a single NVIDIA H200 GPU with 141GB memory. The detailed hyperparameters are reported in Ap- pendix. To identify context-aware neurons, we use approx- imately 400 sub-samples of HotpotQA training data. We set frequency occurrence thresholdkas 130 to ensure 100 neurons are selected for bothP rel andP irrel and 30 neurons forP shared . During data construction of theD RS , the exter- nal LLM produces 142-token summaries grounded exclu- sively in the provided input documents, with zero intrinsic knowledge usage enforced via prompting, thereby ensuring document-grounded generation and mitigating hallucination. Our two-stage instruction tuning both use the AdamW opti- mizer, with the denoising stage trained at a learning rate of 1× 10 −5 for one epoch and the noise filtering stage trained at a learning rate of2× 10 −5 for two epochs, using a batch size of 4 throughout. 4.4. Main Results We analyze the results in Table 1 by comparing three con- figurations: the base model without ranking, the base model 7 https://huggingface.co/naver/trecdl22-crossencoder-debertav3 8 https://huggingface.co/MignonMiyoung/RankCoT 5 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model MethodNQASQATriviaQASCIQPOPQAHotpotQA2WikiAvg. RALM (Lewis et al., 2020)62.2167.8288.6053.3060.8545.4442.2760.07 + Reranker (Glass et al., 2022)61.6567.9388.3152.6060.3745.2741.8459.71 + RankCoT (Wu et al., 2025b)59.4365.9389.6057.4059.8048.5443.4260.59 RetRobust (Yoran et al., 2024)53.2361.9289.8058.2053.2440.8435.1756.06 + Reranker (Glass et al., 2022)54.4264.1390.5558.2054.9241.4635.6957.05 + RankCoT (Wu et al., 2025b)53.2660.5588.9357.8053.1843.8439.9456.79 InstructRAG (Wei et al., 2024)60.6266.9887.4855.3060.2545.3344.6160.08 + Reranker (Glass et al., 2022)63.5871.5190.7356.8064.6048.3247.0163.22 + RankCoT (Wu et al., 2025b)60.1767.7290.7659.5062.2451.6650.0663.16 PA-RAG (Wu et al., 2025a)67.9274.4789.8656.6064.3749.2946.5764.15 + Reranker (Glass et al., 2022)72.0179.4392.5857.6071.0854.1248.8767.96 + RankCoT (Wu et al., 2025b)66.3073.1092.3263.4065.6853.8952.0266.67 Neuro-RIT69.6875.3192.2160.8067.3752.2546.3466.28 + Reranker (Glass et al., 2022)72.5780.4893.91 † 61.6074.29 † 55.71 † 47.8069.48 + RankCoT (Wu et al., 2025b)61.9369.0991.6059.1063.2852.0451.1364.02 Table 1. Main results of performance comparison across QA benchmarks. We report Accuracy for each dataset. The best results are highlighted in bold and the second-best results areunderlined(per metric). † :Neuro-RITsignificantly improves the best baseline at p < 0.05 with paired t-test. RALM RetRobust InstructRAG PA-RAG Neuro-RIT 40 50 60 70 LLM Eval 73.55 61.44 37.89 68.08 76.87 NQ RALM RetRobust InstructRAG PA-RAG Neuro-RIT 40 50 60 70 80 75.85 50.60 36.14 77.59 77.71 ASQA RALM RetRobust InstructRAG PA-RAG Neuro-RIT 74 76 78 80 82 84 86 88 90 87.50 82.55 78.86 86.93 87.98 TriviaQA RALM RetRobust InstructRAG PA-RAG Neuro-RIT 30 35 40 45 50 55 60 57.93 40.64 34.30 59.54 59.81 POPQA RALMRetRobustInstructRAGPA-RAGNeuro-RIT Figure 2. Performance comparison of four RAG-based methods on four QA benchmarks, evaluated using an LLM-based metric. augmented with aReranker, and the base model aug- mented with RankCoT. We first examine the performance of base models without any ranking components. Across all datasets,Neuro-RIT consistently achieves strong performance compared to existing baselines, includingRALM,RetRobust, and InstructRAG, while remaining comparable to or slightly outperformingPA-RAG. Notably, this strong base perfor- mance is observed despiteNeuro-RITbeing trained exclu- sively on HOTPOTQA, indicating that the learned behavior generalizes beyond the training distribution to the remaining six datasets. These results suggest that Neuro-RIT’s neuron- level disentanglement and denoising enable the generator to robustly utilize relevant evidence while suppressing the influence of irrelevant retrieved contexts. When the base model is augmented with aReranker, performance is consistently improved forRetRobust,InstructRAG, PA-RAG, andNeuro-RITacross all benchmarks, with particularly pronounced gains on single-hop datasets such as NQ, ASQA, and TRIVIAQA. Under this configura- tion,Neuro-RITachieves the highest average accuracy of 69.48, establishing a clear improvement over all baseline configurations. In contrast, augmenting the base model with reasoning-aware ranking viaRankCoTyields a different performance pattern: while its impact is limited on single- hop benchmarks, it provides the greatest improvements on 2WIKI, which requires multi-hop reasoning over longer contexts. We further validate these observations under an LLM-based evaluation metric, LLMEval (Rau et al., 2024), on four QA benchmarks. Consistent with the automatic evalua- tion results in Table 1,Neuro-RITachieves the highest LLMEval scores across all four benchmarks, as shown in Figure 2. This suggests that the proposed neuron-level disen- tanglement and denoising improve not only answer accuracy but also the faithfulness and quality of evidence utilization, as corroborated by the LLM-based evaluation. 6 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model 012345678910111213141516171819202122232425262728293031 Layer Index 0 1 2 3 4 5 6 7 8 9 10 11 12 13 Number of Neurons RelSharedIrrel 012345678910111213141516171819202122232425262728293031 Layer Index 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Number of Neurons RelShared+Irrel Figure 3. Neuron distribution of LLaMA-3-8B-Instruct on the HotpotQA dataset; the left panel shows the distribution ofP rel ,P shared and P irrel , while the right panel shows the P rel against the combination of P irrel and P shared . ModelNQASQATriviaQASCIQPOPQAHotpotQA2WikiAVG. RALM64.8571.2089.5960.8062.7848.3948.8063.77 + Reranker68.4873.6292.2661.9069.0550.6249.3866.47 + RankCoT60.3166.7790.4657.8060.8349.7347.9361.98 Neuro-RIT66.4772.5790.0462.5064.1550.4650.6565.26 + Reranker69.7975.4292.6363.6070.6553.2551.0568.06 + RankCoT63.4869.6291.7160.7063.1952.4551.1764.61 Table 2. Comparison of accuracy on Mistral-7B-Instruct-v0.2 across all datasets. Best results are shown in bold and second-best results are underlined. 012345678910111213141516171819202122232425262728293031 Layer Index 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Number of Neurons RelSharedIrrel Figure 4. Neuron distribution of Mistral-7B-Instruct-v0.2 on the HotpotQA dataset showing the distribution ofP rel ,P shared , and P irrel . 4.5. Analysis Layer-wise Distribution of Identified Neurons.Figure 3 (Left) shows the layer-wise distribution of the following set of neurons,P rel ,P irrel andP shared in our generator LLM, LLaMA-3-8B-Instruct(Dubey et al., 2024), that are obtained from the context-aware neuron mining. We observe that P rel is densely populated in the mid layers (approximately 15–19th layers), with additional occurrences in the upper layers (e.g., 29–31st layers), rather than in the lower lay- ers. In contrast,P shared andP irrel are sparsely distributed across early-to-mid layers, with a concentration observed in the higher layers (29–31st layers). This trend visualize distinctly when exploring the distribution ofP irrel together withP shared . Specifically, we show the occurrence ofP irrel together withP shared in contrast toP rel in Figure 3 (Right). WhileP irrel together withP shared are concentrated in the higher layers,P rel are primarily observed in the mid-layers, with continued presence in the upper layers. Overall, we conclude that context-aware neurons shows a distinct layer- wise separation. Parameter Efficiency Compared to Full Fine-tuning. Our approach demonstrates substantial parameter efficiency when compared to full fine-tuning. While full fine-tuning updates approximately8billion parameters, our method selectively updates only about0.529billion parameters, cor- responding to approximately6.6%of the entire model. As a result, we reduce the number of trainable parameters by roughly93.4%, while achieving performance comparable to or even exceeding that of full fine-tuning. These findings indicate that selectively training a carefully chosen subset of neurons that play a critical role in model performance can serve as an effective alternative to full fine-tuning, offering significant reductions in computational cost and memory usage. Generator Substitution with Mistral. We evaluate whether the proposed approach generalizes to a different generator LLM beyond the original LLaMA-based setup. To this end, we replace the generator with Mistral-7B- Instruct-v0.2 (Jiang et al., 2023) while keeping all other components unchanged. As shown in Table 2, under the Mistral-7B-Instruct setting, our approach combined with a Rerankerachieves the highest average accuracy of 68.06. The strongest baseline,RALM+Reranker, attains an aver- age accuracy of 66.47, while all other configurations yield lower performance. Even without reranking, our method achieves an average accuracy of 65.26, outperforming both the vanilla RALM and RALM+RankCoT. In addition, we also examine the layer-wise distribution of selected neurons using identical criteria as in the LLaMA- based setting for Mistral. As illustrated in Figure 4, neurons categorized asP rel are primarily concentrated in the mid- dle layers, with an additional peak in the uppermost layers, whileP shared andP irrel emerge in the middle layers and be- come increasingly prevalent toward the upper layers, closely mirroring the distribution observed for LLaMA. Based on this distribution, we select layers with highP shared andP irrel concentration and fine-tune only the corresponding parame- ters following the same layer selection and adaptation pro- 7 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model RelevantIrrelevant ModelNQHotpotQASCIQNQHotpotQASCIQ RALM79.1673.1571.0820.4315.1917.49 InstructRAG76.8774.2975.0419.5411.9816.62 PA-RAG86.3881.9679.4520.3012.3212.83 Neuro-RIT86.7381.7582.5025.7618.4520.41 Table 3. The comparison of the accuracy of our method in scenar- ios with relevant (answer-present) and irrelevant (answer-absent) retrieved contexts. ModelNQHotpotQATriviaQASCIQPOPQAAvg. Neuro-RIT69.6852.2592.2160.8067.3768.46 w/o denoising67.8850.3791.1656.6065.8766.38 w/o NoiseFilter61.5044.8287.8852.1059.6061.18 w/o NoiseFilter, Layers61.9347.1088.9851.9058.9461.77 w/o NoiseFilter, Neurons69.0153.2691.8559.9067.0068.20 w/o NoiseFilter,D RS 67.5351.6091.1458.1065.8266.84 Table 4. Comparison of performance across various QA datasets under different ablation settings ofNeuro-RIT. Average accu- racy is computed over the five QA datasets. cedure. The consistent performance gains and aligned distri- butional patterns together indicate that the proposed method- ology transfers reliably from LLaMA to Mistral, supporting the view that our approach is generator-agnostic and cap- tures a general mechanism for improving robustness against noisy retrievals for RALMs. Irrelevant Contexts Handling. To verify the robustness ofNeuro-RIT, we evaluate its performance in two distinct scenarios: 1) where retrieved contexts contain the ground- truth answer (Relevant), and 2) where they do not (Irrele- vant). As summarized in Table 3,Neuro-RITconsistently outperforms all baseline methods across both settings. No- tably, in the Irrelevant scenario—where model robustness is most severely tested—Neuro-RITachieves the highest accuracy on all three datasets (NQ, HotpotQA, and SCIQ), significantly surpassingPA-RAG. This suggests that our approach more effectively filters out retrieval noise and miti- gates the risk of hallucination without the need for additional external modules. 4.6. Ablation Study Contribution of Individual Components.To analyze the contribution of individual modules in our framework, we conduct an ablation study by removing each module one at a time. Specifically, we evaluate the performance of our method 1) without the denoising stage (w/o denoising), and 2) without noise filtering enhancement (w/o NoiseFilter), and show results in Table 4. We find that removing the de- noising stage (w/o denoising) consistently degrades perfor- mance, reducing the average accuracy from 68.46 to 66.38 (particularly evident on NQ and SCIQ). We also observe that removing the noise filtering enhancement stage (w/o ActNoiseFilter) leads to an accuracy drop, from 68.46 to 61.18, suggesting the effectiveness of our method. Overall, we observe performance degradation across both settings, which confirms the effectiveness and necessity of each mod- ule in Neuro-RIT. Effect of Noise Filtering Enhancement Components During the noise filtering enhancement instruction tuning, we update two categorization of parameters identified by context-aware neuron mining: 1) layers thatP irrel andP shared are densely populated, and 2) the neurons themselves,P irrel , P shared andP rel . To investigate the impact of each catego- rization, we remove each candidate for instruction tuning, and show results in Table 4. We observe that excluding layers (w/o NoiseFilter, Layers) leads to a substantial per- formance drop, decreasing the average score from 68.46 to 61.77. This effect is consistently observed across all benchmarks, with pronounced degradation on SCIQ and POPQA. In contrast, excluding neurons (w/o NoiseFilter, Layers) results in a slight reduction in the average score, decreasing from 68.46 to 68.20, with observable effects on NQ and SCIQ. We also conduct experiments replacing D RS toD Q (i.e., the question-answer pairs), which leverage ground-truth answers as target outputs. We observe perfor- mance degradation as well, which proves that constructing D RS for relevant evidence distillation. 5. Conclusion In this paper, we proposedNeuro-RIT, a neuron-level instruction tuning framework for improving robustness for retrieval-augmented language models (RALM). Specifically, we explicitly identify relevant and irrelevant context-aware neurons via attribution analysis and enforce robustness through a two-stage instruction tuning strategy that sup- presses the influence of irrelevant evidence while preserv- ing the model’s ability to leverage relevant context. By selectively adapting task-critical neurons and layers, our approach departs from dense fine-tuning paradigms and instead exploits the inherent sparsity of large language models. Extensive experiments across diverse knowledge- intensive and multi-hop question answering benchmarks demonstrate thatNeuro-RITconsistently outperforms strong RALM baselines and recent robustness-enhancing methods, while achieving substantial parameter efficiency. Notably,Neuro-RITgeneralizes well to unseen datasets and different generator architectures despite being trained on a single dataset, indicating that neuron-level robust- ness supervision captures transferable retrieval-aware behav- iors. These results suggest that fine-grained, neuron-level adaptation provides an effective and scalable alternative to full-parameter or layer-level tuning for robust retrieval- augmented generation, and highlight a promising direction for future work on controllable and reliable large language model adaptation. 8 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model References Chen, J., Lin, H., Han, X., and Sun, L. Benchmarking large language models in retrieval-augmented generation. In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, p. 17754–17762, 2024. Cuconasu, F., Trappolini, G., Siciliano, F., Filice, S., Cam- pagnano, C., Maarek, Y., Tonellotto, N., and Silvestri, F. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 719–729, 2024. Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. Fan, W., Ding, Y., Ning, L., Wang, S., Li, H., Yin, D., Chua, T.-S., and Li, Q. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, p. 6491–6501, 2024. Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Sun, J., and Wang, H. Retrieval-augmented generation for large language models: A survey. 2023. Glass, M., Rossiello, G., Chowdhury, M. F. M., Naik, A., Cai, P., and Gliozzo, A. Re2G: Retrieve, rerank, gener- ate. In Carpuat, M., de Marneffe, M.-C., and Meza Ruiz, I. V. (eds.), Proceedings of the 2022 Conference of the North American Chapter of the Association for Compu- tational Linguistics: Human Language Technologies, p. 2701–2715, Seattle, United States, July 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022. naacl-main.194.URLhttps://aclanthology. org/2022.naacl-main.194/. Guu, K., Lee, K., Tung, Z., Pasupat, P., and Chang, M. Retrieval augmented language model pre-training. In International conference on machine learning, p. 3929– 3938. PMLR, 2020. Ho, X., Nguyen, A.-K. D., Sugawara, S., and Aizawa, A. Constructing a multi-hop qa dataset for compre- hensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060, 2020. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxon- omy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1–55, 2025. Izacard, G., Lewis, P., Lomeli, M., Hosseini, L., Petroni, F., Schick, T., Dwivedi-Yu, J., Joulin, A., Riedel, S., and Grave, E. Atlas: Few-shot learning with retrieval aug- mented language models. Journal of Machine Learning Research, 24(251):1–43, 2023. Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., de las Casas, D., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., Lavaud, L. R., Lachaux, M.- A., Stock, P., Scao, T. L., Lavril, T., Wang, T., Lacroix, T., and Sayed, W. E. Mistral 7b, 2023. URLhttps: //arxiv.org/abs/2310.06825. Jiang, Y., Rajendran, G., Ravikumar, P., and Aragam, B. Do llms dream of elephants (when told not to)? latent con- cept association and associative memory in transformers. Advances in Neural Information Processing Systems, 37: 67712–67757, 2024. Joshi, M., Choi, E., Weld, D. S., and Zettlemoyer, L. Trivi- aqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 1601–1611, 2017. Kim, S., Kim, D., Park, C., Lee, W., Song, W., Kim, Y., Kim, H., Kim, Y., Lee, H., Kim, J., Ahn, C., Yang, S., Lee, S., Park, H., Gim, G., Cha, M., Lee, H., and Kim, S. SOLAR 10.7B: Scaling large language models with simple yet effective depth up-scaling. In Yang, Y., Da- vani, A., Sil, A., and Kumar, A. (eds.), Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 6: Industry Track), p. 23–35, Mexico City, Mexico, June 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024. naacl-industry.3. URLhttps://aclanthology. org/2024.naacl-industry.3/. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for ques- tion answering research. Transactions of the Association for Computational Linguistics, 7:453–466, 2019. Lassance, C., D ́ ejean, H., Formal, T., and Clinchant, S. Splade-v3: New baselines for splade, 2024.URL https://arxiv.org/abs/2403.06789. Leng, Y. and Xiong, D. Towards understanding multi-task learning (generalization) of llms via detecting and ex- ploring task-specific neurons. In Proceedings of the 31st International Conference on Computational Linguistics, p. 2969–2987, 2025. 9 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K ̈ uttler, H., Lewis, M., Yih, W.-t., Rockt ̈ aschel, T., et al. Retrieval-augmented generation for knowledge- intensive nlp tasks. Advances in neural information pro- cessing systems, 33:9459–9474, 2020. Li, Z., You, C., Bhojanapalli, S., Li, D., Rawat, A. S., Reddi, S. J., Ye, K., Chern, F., Yu, F., Guo, R., et al. The lazy neuron phenomenon: On emergence of activation sparsity in transformers. arXiv preprint arXiv:2210.06313, 2022. Lin, W., Gerchanovsky, A., Akgul, O., Bauer, L., Fredrik- son, M., and Wang, Z. Llm whisperer: An inconspicuous attack to bias llm responses. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Sys- tems, p. 1–24, 2025. Lin, X. V., Chen, X., Chen, M., Shi, W., Lomeli, M., James, R., Rodriguez, P., Kahn, J., Szilvasy, G., Lewis, M., et al. Ra-dit: Retrieval-augmented dual instruction tuning. In The Twelfth International Conference on Learning Repre- sentations, 2023. Mallen, A., Asai, A., Zhong, V., Das, R., Khashabi, D., and Hajishirzi, H. When not to trust language models: Inves- tigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 9802–9822, 2023. Rau, D., D ́ ejean, H., Chirkova, N., Formal, T., Wang, S., Clinchant, S., and Nikoulina, V. Bergen: A benchmarking library for retrieval-augmented generation. In Findings of the Association for Computational Linguistics: EMNLP 2024, p. 7640–7663, 2024. Shi, D., Jin, R., Shen, T., Dong, W., Wu, X., and Xiong, D. Ircan: Mitigating knowledge conflicts in llm generation via identifying and reweighting context-aware neurons. Advances in Neural Information Processing Systems, 37: 4997–5024, 2024. Stelmakh, I., Luan, Y., Dhingra, B., and Chang, M.-W. ASQA: Factoid questions meet long-form answers. In Goldberg, Y., Kozareva, Z., and Zhang, Y. (eds.), Proceed- ings of the 2022 Conference on Empirical Methods in Nat- ural Language Processing, p. 8273–8288, Abu Dhabi, United Arab Emirates, December 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022. emnlp-main.566. URLhttps://aclanthology. org/2022.emnlp-main.566/. Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribu- tion for deep networks. In International conference on machine learning, p. 3319–3328. PMLR, 2017. Trivedi, H., Balasubramanian, N., Khot, T., and Sabharwal, A. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 10014–10037, 2023. Vig, J., Fabbri, A., Kryscinski, W., Wu, C.-S., and Liu, W. Exploring neural models for query-focused summariza- tion. In Carpuat, M., de Marneffe, M.-C., and Meza Ruiz, I. V. (eds.), Findings of the Association for Computa- tional Linguistics: NAACL 2022, p. 1455–1468, Seat- tle, United States, July 2022. Association for Computa- tional Linguistics. doi: 10.18653/v1/2022.findings-naacl. 109. URLhttps://aclanthology.org/2022. findings-naacl.109/. Wang, S. and Yu, Y. iQUEST: An iterative question-guided framework for knowledge base question answering. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 15616–15628, Vienna, Austria, July 2025. Association for Computational Linguistics. ISBN 979-8-89176-251-0. URLhttps://aclanthology. org/2025.acl-long.760/. Wang, Y., Li, P., Sun, M., and Liu, Y. Self-knowledge guided retrieval augmentation for large language mod- els.In Bouamor, H., Pino, J., and Bali, K. (eds.), Findings of the Association for Computational Lin- guistics:EMNLP 2023, p. 10303–10315, Singa- pore, December 2023. Association for Computational Linguistics.doi: 10.18653/v1/2023.findings-emnlp. 691. URLhttps://aclanthology.org/2023. findings-emnlp.691/. Wang, Z., Liu, A., Lin, H., Li, J., Ma, X., and Liang, Y. Rat: Retrieval augmented thoughts elicit context-aware reasoning in long-horizon generation. arXiv preprint arXiv:2403.05313, 2024. Wei, Z., Chen, W.-L., and Meng, Y. Instructrag: Instruct- ing retrieval-augmented generation via self-synthesized rationales. arXiv preprint arXiv:2406.13629, 2024. Welbl, J., Liu, N. F., and Gardner, M. Crowdsourcing multiple choice science questions. In Proceedings of the 3rd Workshop on Noisy User-generated Text, p. 94–106, 2017. Wu, J., Cai, H., Yan, L., Sun, H., Li, X., Wang, S., Yin, D., and Gao, M. PA-RAG: RAG alignment via multi-perspective preference optimization. In Chiruzzo, L., Ritter, A., and Wang, L. (eds.), Proceedings of the 2025 Conference of the Nations of the Americas Chap- ter of the Association for Computational Linguistics: 10 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model Human Language Technologies (Volume 1: Long Pa- pers), p. 9091–9112, Albuquerque, New Mexico, April 2025a. Association for Computational Linguistics. ISBN 979-8-89176-189-6. doi: 10.18653/v1/2025.naacl-long. 459. URLhttps://aclanthology.org/2025. naacl-long.459/. Wu, M., Liu, Z., Yan, Y., Li, X., Yu, S., Zeng, Z., Gu, Y., and Yu, G. RankCoT: Refining knowledge for retrieval- augmented generation through ranking chain-of-thoughts. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 12857–12874, Vienna, Austria, July 2025b. Association for Computational Linguistics. ISBN 979-8-89176-251-0. URLhttps://aclanthology. org/2025.acl-long.629/. Xu, F., Shi, W., and Choi, E. Recomp: Improving retrieval- augmented lms with compression and selective augmen- tation. arXiv preprint arXiv:2310.04408, 2023. Xu, H., Zhan, R., Ma, Y., Wong, D. F., and Chao, L. S. Let’s focus on neuron: Neuron-level supervised fine-tuning for large language model. In Proceedings of the 31st International Conference on Computational Linguistics, p. 9393–9406, 2025. Yang, Z., Qi, P., Zhang, S., Bengio, Y., Cohen, W., Salakhut- dinov, R., and Manning, C. D. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Meth- ods in Natural Language Processing, p. 2369–2380, 2018. Yoran, O., Wolfson, T., Ram, O., and Berant, J. Making retrieval-augmented language models robust to irrelevant context. In The Twelfth International Conference on Learning Representations (ICLR), 2024. Yu, W., Zhu, C., Zhang, Z., Wang, S., Zhang, Z., Fang, Y., and Jiang, M. Retrieval augmentation for common- sense reasoning: A unified approach. In Goldberg, Y., Kozareva, Z., and Zhang, Y. (eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Lan- guage Processing, p. 4364–4377, Abu Dhabi, United Arab Emirates, December 2022. Association for Compu- tational Linguistics. doi: 10.18653/v1/2022.emnlp-main. 294. URLhttps://aclanthology.org/2022. emnlp-main.294/. 11 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model A. Supplementary Materials A.1. Data for Neuron Attribution To compute neuron attribution scores, we adopt the Integrated Gradients (IG) formulation (Sundararajan et al., 2017) and construct a dedicated binary decision dataset tailored for attribution analysis. The dataset is designed to probe neuron behaviors under contrasting retrieval conditions by explicitly controlling the relevance of the provided context. Data Construction for Neuron AttributionWe employ a standard RAG pipeline with the SPLADE-v3(Lassance et al., 2024) retriever over the HOTPOTQA dataset(Yang et al., 2018) For each query, SPLADE-v3 assigns sparse lexical similarity scores to candidate documents. Based on these scores, we construct two attribution subsets: a ̃ P irrel and ̃ P rel , each consisting of 389 instances. Due to the substantial memory footprint of Integrated Gradients at the neuron level, each attribution subset is limited to 389 instances to ensure feasible computation on a single NVIDIA H200 GPU. The ̃ P rel attribution set is formed by selecting the 389 highest-scoring retrieved contexts (top-1 per query), while the ̃ P irrel attribution set is constructed by selecting the 389 lowest-scoring contexts. In cases where multiple contexts receive zero similarity scores, instances are selected according to their retrieval order. The retrieved contexts are used in their raw form without additional filtering or summarization, and all attribution instances are constructed through deterministic preprocessing using Python scripts. Binary Attribution Task FormulationEach instance is formulated as a forced binary classification task. Specifically, the model is provided with a single combined input consisting of a question, a retrieved context, and a proposed answer. The model is instructed to determine whether the proposed answer can be derived solely from the given context by choosing between two predefined options:YESorNO. This formulation enforces a strict binary decision and eliminates ambiguity in the output space, enabling stable and interpretable attribution analysis. An illustrative example of the attribution dataset format is provided in Table 5. Attribution scores are computed with respect to the loss associated with predicting the correct binary label. During IG computation, the model is restricted to inferring answer correctness based solely on the provided context, without access to any external knowledge. The resulting IG scores quantify the contribution of individual neurons to the model’s binary decision and are subsequently used to identify neurons with distinct functional roles. A.2. Examples of the Constructed Datasets We present qualitative examples of the two instruction-tuning datasets constructed in our framework: (i) an irrelevant-context denoising dataset used to suppress irrelevant context-aware neurons, and (i) a relevant summary datasetD RS used to enforce evidence distillation from relevant contexts. Irrelevant Context Denoising Dataset.For the denoising stage, we construct training instances by selecting five retrieved documents that are irrelevant to the input query. Given these irrelevant contexts, the model is explicitly instructed to identify the absence of query-relevant information and emit an End-of-Text (EOT) token. This dataset is designed to impose a hard constraint on neurons associated with irrelevant contexts, effectively decoupling their activations from the generation process. As a result, the model learns to terminate generation when only non-informative or misleading evidence is provided. A representative example of this denoising dataset is shown in Table 6. Relevant Summary DatasetD RS .In contrast, the second dataset focuses on reinforcing the extraction of query-relevant evidence. For each query, we retrieve five candidate documents and employ an external LLM to construct a concise relevant summary that integrates only the information semantically aligned with the query. For example, given the query “Who was once considered the best kickboxer in the world, however he has been involved in a number of controversies relating to his unsportsmanlike conduct and crimes of violence outside of the ring?”, the generated summary correctly identifies Badr Hari as the target entity. Among the retrieved documents, only the passage describing Badr Hari contains both the required biographical information (e.g., “Moroccan-Dutch super heavyweight kickboxer,” “former K-1 Heavyweight champion”) and contextual cues regarding unsportsmanlike conduct and violent incidents. Other retrieved documents referring to unrelated athletes (e.g., Rau’shee Warren, Yamaguchi Falc ̃ ao, or Rafael Carvalho) are excluded from the summary due to a lack of semantic correspondence to the query. An illustrative example of the constructed relevant summary is provided in Table 7. These examples show thatD RS is constructed to focus on selective evidence integration, rather than holistic document summarization. Together, the two datasets form a complementary training signal: the first enforces strict suppression of 12 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model irrelevant contexts via EOT generation, while the second guides the model to distill and summarize only query-relevant information. A.3. Prompt We provide the prompts used for Denoising Irrelevant Context-Aware Neuron-level Parameters and Context-Aware Neuron- level Parameters Robustness Enforcing Fine-tuning in Section 3.2. Prompt for Denoising Context-Aware Neuron-level Parameters <|beginoftext|><|startheaderid|>user<|endheaderid|> Given a document and a query, reason step by step to identify only the parts of the document that are directly relevant to the query, and provide a concise summary of those relevant parts. Background: Document 1: document 1 Document 2: document 2 Document 3: document 3 Document 4: document 4 Document 5: document 5 Question: question<|eotid|><|startheaderid|>assistant<|endheaderid|><|eotid|> Prompt for Context-Aware Neuron-level Parameters Robustness Enforcing Fine-tuning <|beginoftext|><|startheaderid|>user<|endheaderid|> Given a document and a query, reason step by step to identify only the parts of the document that are directly relevant to the query, and provide a concise summary of those relevant parts. Background: Document 1: document 1 Document 2: document 2 Document 3: document 3 Document 4: document 4 Document 5: document 5 Question: question<|eot id|><|startheaderid|>assistant<|endheaderid|> relevant summary<|eotid|> Prompt for Relevant Summary Data Construction Below is a document. Your task is to find and concisely summarize only the parts of the document that are directly relevant to the given query. - Do not summarize the entire document. - Exclude any information that is not related to the query. - Focus only on the key points that are most relevant to the query. Query: query Document: document [Relevant Summary] (For this section, reason step-by-step in a Chain-of-Thought (CoT) manner to identify the relevant information. Show your thinking process as you determine which parts of the document are relevant to the query. Then, Summarize only the information directly related to the query based on your reasoning. If nothing is relevant, leave this section blank.) 13 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model Prompt for Open-Domain Question Answering <|beginoftext|><|startheaderid|>system<|endheaderid|> You are a helpful assistant. Your task is to extract relevant information from the provided documents and answer questions as briefly as possible.<|eot id|> <|start headerid|>user<|endheaderid|> Background: Document 1: document 1 Document 2: document 2 Document 3: document 3 Document 4: document 4 Document 5: document 5 Question: question<|eot id|><|startheaderid|>assistant<|endheaderid|> A.4. Evaluation Evaluation with Match MetricAll evaluations are conducted using the BERGEN framework (Rau et al., 2024) under a retrieval-augmented generation (RAG) setting. We evaluate our method on the official test splits of the KILT-HotpotQA (Yang et al., 2018), and assess generalization across multiple benchmarks, including KILT-NQ (Kwiatkowski et al., 2019), ASQA (Stelmakh et al., 2022), KILT-TriviaQA (Joshi et al., 2017), SCIQ (Welbl et al., 2017), POPQA (Mallen et al., 2023) and 2WikiMultiHopQA (Ho et al., 2020). No evaluation instances are used during training or neuron attribution computation. For all experiments, we fix the retriever to SPLADE-v3 to ensure consistent retrieval behavior across methods. Given an input question, the retriever first retrieves the top-50 candidate documents based on sparse lexical similarity, from which the top-5 documents are provided to the generator as contextual input. To ensure fair comparison, a unified prompt template is applied across all models and datasets. Model performance is evaluated using the Match metric provided by the BERGEN framework, which measures exact correspondence between the generated response and the ground-truth answer. All Match-based evaluations are conducted with a batch size of 256 for all models. LLM-based Evaluation Protocol In addition to exact match evaluation, we conduct LLM-based evaluations using the BERGEN framework to assess answer quality beyond strict string matching. LLM-based evaluation is performed on four representative QA benchmarks: NQ, ASQA, TriviaQA, and POPQA. For these datasets, we evaluate our method and baseline approaches on the development splits. Following the BERGEN evaluation protocol, we useSOLAR-10.7B(Kim et al., 2024) as the evaluation model, deployed via vLLM. The LLM evaluator is configured with a batch size of 128. All evaluated methods use identical retrieved contexts and prompts to ensure a controlled comparison. 14 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model Examples of Relevant Attribution Set ( ̃ P rel ) cContext: Arthur’s Lady’s Home Magazine. Arthur’s Lady’s Home Magazine Arthur’s Home Magazine (1852–ca.1898) or Ladies’ Home Magazine was an American periodical published in Philadelphia by Timothy Shay Arthur. Editors Arthur and Virginia Francis Townsend selected writing and illustrations intended to appeal to female readers. Among the contributors: Mary Tyler Peabody Mann and Kate Sutherland. In its early years the monthly comprised a selection of articles originally published in Arthur’s weekly “Home Gazette.” Its nonfiction stories contained occasional factual inaccuracies for the sake of a good read.→ Question: Which magazine was started first Arthur’s Magazine or First for Women? Proposed Answer: Arthur’s Magazine qIf the proposed answer can be derived by referring to the context, answer YES; otherwise, answer NO. The correct answer is choices[’YES’, ’NO’] gold answer0 cContext: The Oberoi Group. The Oberoi Group is a hotel group with its head office in Delhi. Founded in 1934, the company owns and/or operates 35 luxury hotels and two river cruise ships in six countries, primarily under its Oberoi Hotels & Resorts and Trident Hotels brands. The foundations of the Oberoi Group date back to 1934 when Rai Bahadur Mohan Singh Oberoi bought two properties in Delhi and Shimla.→ Question: The Oberoi family is part of a hotel company that has a head office in what city? Proposed Answer: Delhi qIf the proposed answer can be derived by referring to the context, answer YES; otherwise, answer NO. The correct answer is choices[’YES’, ’NO’] gold answer0 Examples of Irrelevant Attribution Set ( ̃ P irrel ) cContext: Autism. Sustained special education programs and behavior therapy early in life can help children acquire self-care, communication, and job skills, and often improve functioning and decrease symptom severity and maladaptive behaviors. Educational interventions often used include applied behavior analysis (ABA), speech and language therapy, social skills therapy, and occupational therapy.→ Question: Which magazine was started first Arthur’s Magazine or First for Women? Proposed Answer: Arthur’s Magazine qIf the proposed answer can be derived by referring to the context, answer YES; otherwise, answer NO. The correct answer is choices[’YES’, ’NO’] gold answer0 cContext: Albedo. Albedo is the measure of the diffuse reflection of solar radiation out of the total solar radiation received by an astronomical body. Enceladus has one of the highest known albedos in the Solar System, reflecting about 99% of incident radiation.→ Question: The Oberoi family is part of a hotel company that has a head office in what city? Proposed Answer: Delhi qIf the proposed answer can be derived by referring to the context, answer YES; otherwise, answer NO. The correct answer is choices[’YES’, ’NO’] gold answer0 Table 5. Case I: Representative examples of the binary decision dataset used for Integrated Gradients-based neuron attribution. The dataset frames the task as a forced binary choice (YES/NO) to determine if the answer is strictly derivable from the context. This allows us to disentangle neurons responsible for processing relevant contexts ( ̃ P rel ) from those processing irrelevant ones ( ̃ P irrel ). 15 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model QuestionWhich magazine was started first Arthur’s Magazine or First for Women? Document 1Autism. sustained special education programs and behavior therapy early in life can help children acquire self-care, communication, and job skills, and often improve functioning and decrease symptom severity and maladaptive behaviors; claims that intervention by around age three years is crucial are not substantiated. While medications have not been found to help with core symptoms, they may be used for associated symptoms, such as irritability, inattention, or repetitive behavior patterns. Section:Management.:Education. Educational interventions often used include applied behavior analysis (ABA), developmental models, structured teaching, speech and language therapy, social skills therapy, and occupational therapy. Among these approaches, interventions either treat autistic Document 2Albedo. back into space than what they absorb, effectively cooling the Earth. This has been a concern since arctic ice and snow has been melting at higher rates due to higher temperatures, creating regions in the arctic that are notably darker (being water or ground which is darker color) and reflects less heat back into space. This feedback loop results in a reduced albedo effect. Section:Examples of terrestrial albedo effects.:Climate and weather. Albedo affects climate by determining how much radiation a planet absorbs. The uneven heating of Earth from albedo variations between land, ice, or ocean surfaces can drive weather. Section:Examples Document 3Autism. social-communication skills in young children, although there is less evidence in its treatment of global symptoms. Neuropsychological reports are often poorly communicated to educators, resulting in a gap between what a report recommends and what education is provided. It is not known whether treatment programs for children lead to significant improvements after the children grow up, and the limited research on the effectiveness of adult residential programs shows mixed results. The appropriateness of including children with varying severity of autism spectrum disorders in the general education population is a subject of current debate among educators and researchers. Section:Management.:Medication. Medications may Document 4Albedo. the mean temperature of the planet would drop to about 0 °C. In contrast, if the entire Earth was covered by water — a so-called ocean planet — the average temperature on the planet would rise to almost 27 °C. Sec- tion:Terrestrial albedo.:White-sky, black-sky, and blue-sky albedo. For land surfaces, it has been shown that the albedo at a particular solar zenith angle “θ” can be approximated by the proportionate sum of two terms: with formula3 being the proportion of direct radiation from a given solar angle, and formula4 being the proportion of diffuse illumination, the actual albedo formula5 (also called Document 5Autism. and the educational system are the main resources for treatment. Services should be carried out by behavior analysts, special education teachers, speech pathologists, and licensed psychologists. Studies of interventions have methodological problems that prevent definitive conclusions about efficacy. However, the development of evidence-based interventions has advanced in recent years. Although many psychosocial interventions have some positive evidence, suggesting that some form of treatment is preferable to no treatment, the methodological quality of systematic reviews of these studies has generally been poor, their clinical results are mostly tentative, and there is little evidence for the relative effectiveness of treatment options. Intensive, Relevant summaryEOT(End of Text) Table 6. Case I: An example of the Irrelevant Context Denoising Dataset. When provided with retrieved documents that are entirely irrelevant to the query, the model is trained to output an End-of-Text (EOT) token, enforcing the suppression of noise. 16 Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model QuestionWhich magazine was started first Arthur’s Magazine or First for Women? Document 1Arthur’s Lady’s Home Magazine. Arthur’s Lady’s Home Magazine Arthur’s Home Magazine (1852-ca.1898) or Ladies’ Home Magazine was an American periodical published in Philadelphia by Timothy Shay Arthur. Editors Arthur and Virginia Francis Townsend selected writing and illustrations intended to appeal to female readers. Among the contributors: Mary Tyler Peabody Mann and Kate Sutherland. In its early years the monthly comprised a selection of articles originally published in Arthur’s weekly “Home Gazette.” Its nonfiction stories contained occasional factual inaccuracies for the sake of a good read. A contemporary review judged it “gotten up in good taste and well; and is in nothing overdone. Even its Document 2Arthur’s Magazine. Arthur’s Magazine Arthur’s Magazine (1844–1846) was an American literary periodical published in Philadelphia in the 19th century. Edited by Timothy Shay Arthur, it featured work by Edgar A. Poe, J.H. Ingraham, Sarah Josepha Hale, Thomas G. Spear, and others. In May 1846 it was merged into “Godey’s Lady’s Book”. A few years later Arthur would launch a new publication entitled “Arthur’s Home Magazine”. Section:Further reading. Section:External links. Document 3Arthur (magazine). Arthur (magazine) Arthur magazine was a bi-monthly periodical that was founded in October 2002, by publisher Laris Kreslins and editor Jay Babcock. It received favorable attention from other periodicals such as “L.A. Weekly”, “Print”, “Punk Planet” and “Rolling Stone”. “Arthur” featured photography and artwork from Spike Jonze, Art Spiegelman, Susannah Breslin, Gary Panter and Godspeed You! Black Emperor. Arthur’s regular columnists included Byron Coley, Thurston Moore, Daniel Pinchbeck, Paul Cullum, Douglas Rushkoff, and T-Model Ford. “Arthur” magazine was particularly drawn to noise music, stoner metal, folk and other types of psychedelia. The first issue of “Arthur” featured an interview with Document 4Marianne Ehrmann. stopped the further issues. Section:Life and career.:Amaliens Erholungsstunden. “Amaliens Erholungsstunden” was the first women’s magazine, and it was self-published monthly by Marianne Ehrmann under her own name. She sent advance notices to acquaintances, friends and potential patrons to distribute these in the circle, and the first six editions resulted in an increase of subscribers. The debts contracted by the self-publishing were indeed paid off, but in all a financial disaster, that’s why in January 1791 Marianne and her husband started a cooperation with the publishing house “J. G. Cottaische Verlagsbuchhandlung” in T ̈ ubingen. Marianne Ehrmann thought to concentrate on the journalistic Document 5Arthur Jewitt. “The Northern Star, or Yorkshire Magazine,” a monthly register of arts, biography, statistics, manufac- tures, &c., which ran to three volumes, 1817–18. On 1 January 1818 he brought out the first number of “The Sylph, or Lady’s Magazine for Yorkshire, Derbyshire, and the adjoining Counties.” “The Lincoln and Lincolnshire Cabinet and Annual Intelligencer” appeared at Lincoln during 1827–9. His “Matlock Companion,” 1835, and “Derbyshire Gems” were very popular. His “Handbook of Practical Perspective,” 1840, and his “Handbook of Geometry,” 1842, were adopted by the committee of council on education. He contributed mathematical papers to the “British Diary” and to the “Lady’s” Relevant summaryThe document mentions “Arthur’s Magazine” as an American literary periodical published from 1844 to 1846. It was edited by Timothy Shay Arthur and later merged into “Godey’s Lady’s Book” in 1846. It also references “Arthur’s Home Magazine” starting in 1852, but no start date or details are given for “First for Women.” There is no mention of “First for Women” magazine’s start date or any comparison to Arthur’s Magazine. Table 7. Case I: An example of the Relevant Summary Dataset (D RS ). The external LLM synthesizes a concise summary using only the query-aligned documents, while effectively filtering out unrelated information. 17