Can I read NeST: Neuron Selective Tuning For LLM Safety on EtoBox?

NeST: Neuron Selective Tuning For LLM Safety by koiredblack is a document available to read on EtoBox.

What is NeST: Neuron Selective Tuning For LLM Safety about?

The document introduces NeST, a novel neuron-structured safety alignment framework for large language models (LLMs) that selectively adapts a small subset of safety-relevant neurons while freezing the rest of the model. NeST achieves significant safety improvements, reducing the average attack success rate from 44.5% to 4.36% with minimal parameter updates, making it highly efficient compared to existing methods. The approach focuses on internal safety mechanisms, enabling targeted adaptations that enhance

Author

koiredblack

Language

EN