Tag
Introduces BUMP, a self-supervised framework for training a profile generator for LLM personalization without task labels, using bidirectional in-batch ranking and GRPO. It matches or outperforms supervised methods on the LaMP benchmark.
The author developed a portable user preference profile system that integrates with ElevenLabs and Pipecat agents, allowing voice assistants to remember user styles and interests across different platforms to skip redundant onboarding.