Donghyun Kim

Personalized visual memory requires fundamentally different reasoning strategies than standard long-context text, as shown by the camroll-agent's superior performance in VQA tasks.

Single-vector embedding models are secretly powerful multi-vector retrievers, and SMART unlocks this hidden potential with a simple plug-and-play upgrade.

Privacy constraints don't just hinder multi-agent collaboration—they fundamentally alter the dynamics, making outcomes more dependent on who initiates the interaction.