RESEARCH · RESEARCH · #2193
FINDERS: routing diffusion states for scenario-adaptive user encoders
This arXiv paper (arXiv:2610.22400v1) proposes FINDERS, a router-equipped framework that repurposes diffusion language models (dLLMs) as a multi-granularity reconstruction space for user representation. FINDERS applies four operations—denoise, factorize, route, and compress—uses learnable meta-slots and a scenario-guided router, and is trained with a new BehaviorQA corpus; the authors report extensive experiments and online A/B tests across ten industrial scenarios in which FINDERS outperforms prior methods in unified and scenario-adaptive settings.
KEY POINTS
- This arXiv paper (arXiv:2610.22400v1) proposes FINDERS, a router-equipped framework that repurposes diffusion language models (dLLMs) as a multi-granularity reconstruction space for user representation.
- FINDERS applies four operations—denoise, factorize, route, and compress—uses learnable meta-slots and a scenario-guided router, and is trained with a new BehaviorQA corpus; the authors report extensive experiments and online A/B tests across ten industrial scenarios in which FINDERS outperforms prior methods in unified and scenario-adaptive settings.
- By enabling scenario-guided, multi-granularity user representations using diffusion LLMs, FINDERS could improve cross-modal personalization and make user encoders more adaptable to diverse downstream tasks.
WHY IT MATTERS
By enabling scenario-guided, multi-granularity user representations using diffusion LLMs, FINDERS could improve cross-modal personalization and make user encoders more adaptable to diverse downstream tasks.