Cognitive-Inspired Multi-Layered Knowledge Suppression Framework for Language Models

This research proposes a novel 'Cognitive Suppression Framework' for aligned language models that operates on three complementary levels: (1) Attentional Filtering Layer, which uses learned attention patterns to suppress harmful content during early transformer layers; (2) Representational Re-encoding Layer, which transforms representations of potentially harmful knowledge into less accessible forms inspired by memory reconsolidation; and (3) Contextual Regulation Layer, which implements flexible suppression adapting to context, recognizing that some knowledge may be harmful in certain situations but beneficial in others. This approach fundamentally differs from current methods by actively restructuring how harmful knowledge is encoded and processed throughout the model architecture, addressing the root cause of persistent harmful representations and aiming for more robust and less brittle AI alignment.

References:

1. ReFT: Representation Finetuning for Language Models. Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Daniel Jurafsky, Christopher D. Manning, Christopher Potts (2024). Neural Information Processing Systems.
2. Linearly Decoding Refused Knowledge in Aligned Language Models. Aryan Shrivastava, Ari Holtzman (2025). arXiv.org.
3. Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models. Guobin Shen, Dongcheng Zhao, Yiting Dong, Xiang He, Yi Zeng (2024). International Conference on Learning Representations.
4. Efficient Jailbreaking of Large Models by Freeze Training: Lower Layers Exhibit Greater Sensitivity to Harmful Content. Hongyuan Shen, Min Zheng, Jincheng Wang, Yang Zhao (2025). arXiv.org.