SHIFT reformulates neuron-level modification as a learnable gate modulation, allowing LLMs to adaptively regulate internal activations for knowledge conflict resolution. Technically, SHIFT equips LLMs with a lightweight gate module and optimizes fewer than 0.01% trainable parameters while keeping the backbone model frozen. During generation, the gate module adjusts the model's internal representations to adaptively leverage contextual and parametric knowledge.
Please refer to the official
GitHub Repository for detailed environment setup, training, and evaluation scripts.
1@misc{li2026shiftgatemodulatedactivationsteering,
2 title={SHIFT: Gate-Modulated Activation Steering for Knowledge Conflict Mitigation in Retrieval-Augmented Generation},
3 author={Ruochang Li and Pengcheng Huang and Zhenghao Liu and Yukun Yan and Huiyuan Xie and Yu Gu and Ge Yu and Maosong Sun},
4 year={2026},
5 eprint={2606.27786},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2606.27786},
9}
