Title: Achieving Controllable Value Alignment in Large Language Models via Neuron-Level Editing

As the impact of large language models (LLMs) on human decision-making and behavior continues to grow, aligning these systems with human values has emerged as a critical priority. Despite this need, current methods that rely on steering for alignment often lack sufficient control; directing a model toward a specific value frequently results in the unintended activation of unrelated values. To better understand this constraint, we introduce the concept of "value leakage," a diagnostic framework that measures the accidental triggering of non-target values during steering. We also propose a normalized leakage metric based on Schwartz’s value theory.

Addressing these challenges, we present NeVA, a novel framework for neuron-level editing designed to enable precise value alignment in LLMs. NeVA operates by pinpointing sparse, value-specific neurons and applying activation editing during inference. This approach allows for fine-grained control without the need for retraining or parameter updates. Our experimental results demonstrate that NeVA not only enhances alignment with target values but also minimizes the negative impact on general model capabilities. Furthermore, the method substantially lowers average leakage, with any remaining effects primarily restricted to value classes that are semantically linked. Ultimately, NeVA provides a more transparent and controllable solution for aligning LLMs with human values.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC