FINDERS Uses Diffusion Language Models for Adaptive User Representation
October 19, 2026
FINDERS replaces static autoregressive user encoders with a router-equipped diffusion language model (dLLM) framework. By utilizing the iterative denoising process, the system factorizes and routes multi-granularity denoising states to better accommodate diverse downstream scenario requirements.
HOW THIS AFFECTS YOU
●
builderThis approach could improve how your system represents complex, multi-modal user behavior for downstream tasks.
●
researcherThis explores using the denoising trajectories of dLLMs as a rich reconstruction space for user modeling.