Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces Neuro-RIT, a neuron-guided instruction tuning framework to improve the robustness of retrieval-augmented language models (RALMs) against irrelevant retrieved contexts. Neuro-RIT uses attribution-based neuron mining to identify neurons responsible for processing relevant vs. irrelevant contexts, then employs a two-stage instruction tuning strategy. Experiments on QA benchmarks show Neuro-RIT outperforms existing robustness-enhancing methods by functionally deactivating noise-specific neurons and optimizing targeted layers for evidence distillation.
RALMs can now handle noisy retrieved contexts far more effectively by surgically deactivating specific neurons trained to process irrelevant information.
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.