RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs
The episode discusses RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs. Hosts discuss how this method enables modular, stackable safety guardrails, allowing developers to apply specialized constraints like those for bias or misinformation independently. The discussion concludes that this approach reframes safety as a core engineering challenge and provides a blueprint for building robust, adaptable AI systems.