SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging

Name: SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging
Uploaded: 2026-08-20T16:59:04.430516+00:00
Duration: 648 s
Description: The episode discusses 'SafeMERGE,' a method for preserving safety alignment when fine-tuning large language models. The hosts explain that SafeMERGE selectively merges layers from a safe model into an unsafe, task-specific model. This approach maintains high utility while significantly reducing harmfulness, offering a practical solution for AI safety.

The episode discusses 'SafeMERGE,' a method for preserving safety alignment when fine-tuning large language models. The hosts explain that SafeMERGE selectively merges layers from a safe model into an unsafe, task-specific model. This approach maintains high utility while significantly reducing harmfulness, offering a practical solution for AI safety.

Open the full episode