Rethinking User Encoders: Routing Diffusion States for Scenario-Adaptive User Representation

FINDERS (FactorIzatioN of Diffusion statEs for Routing across Scenarios) is a router-equipped diffusion language model framework that uses the iterative denoising process as a multi-granularity reconstruction space for scenario-adaptive user representation. The framework comprises four operations—denoise, factorize, route, and compress—and uses learnable meta-slots to organize complementary denoising states into transferable cross-modal semantics and modality-specific behavioral factors. The authors also construct BehaviorQA, a large-scale corpus containing explicit behavioral rationales to provide grounded supervision.