NeuronTune: Fine-Grained Neuron Modulation for Balanced Safety-Utility Alignment in LLMs

2025-08-13

AI Summary

Addressing the longstanding trade-off between safety and utility in large language models (LLMs)—where existing methods often suffer from insufficient safety, excessive rejection of benign queries, or degradation in generation quality and general-task performance—this paper proposes NeuronTune. NeuronTune introduces the first fine-grained, cross-layer sparse neuron modulation framework, overcoming the limitations of conventional layer-wise interventions. It identifies safety-critical neurons via attribution analysis and employs a meta-learning mechanism to adaptively calibrate their activation strength, enabling flexible switching between safety-priority and utility-priority regimes. Experiments demonstrate that NeuronTune significantly enhances robustness against adversarial attacks and improves content safety, while preserving text fluency, factual consistency, and performance on standard benchmarks (e.g., MMLU, BBH) with negligible degradation. It consistently outperforms current state-of-the-art approaches across all evaluated dimensions.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Ensuring robust safety alignment while preserving utility is critical for the reliable deployment of Large Language Models (LLMs). However, current techniques fundamentally suffer from intertwined deficiencies: insufficient robustness against malicious attacks, frequent refusal of benign queries, degradation in generated text quality and general task performance--the former two reflecting deficits in robust safety and the latter constituting utility impairment. We trace these limitations to the coarse-grained layer-wise interventions in existing methods. To resolve this, we propose NeuronTune, a fine-grained framework that dynamically modulates sparse neurons to achieve simultaneous safety-utility optimization. Our approach first identifies safety-critical and utility-preserving neurons across all layers via attribution, then employs meta-learning to adaptively amplify safety-neuron activations and suppress utility-neuron activations. Crucially, NeuronTune enables tunable adjustment of intervention scope via neuron-count thresholds, supporting flexible adaptation to security-critical or utility-priority scenarios. Extensive experimental results demonstrate that our method significantly outperforms existing state-of-the-art technologies, achieving superior model safety while maintaining excellent utility.

Problem

• Balancing safety and utility in Large Language Models deployment
• Addressing coarse-grained layer-wise intervention limitations in existing methods
• Optimizing simultaneous safety-utility via fine-grained neuron modulation

Innovation

• Fine-grained neuron modulation for safety-utility balance
• Meta-learning adaptively adjusts neuron activations
• Tunable intervention scope via neuron-count thresholds