Existing large language model (LLM) based user encoders serialize heterogeneous behavioral logs and compress them into static representations via autoregressive (AR) prediction. The unidirectional constraint restricts bidirectional information flow, hindering cross‑modal interaction and joint preservation of multi‑source signals. Consequently, the learned static representations struggle to satisfy the distinct evidence requirements of diverse downstream scenarios.
We unlock the iterative denoising process of diffusion language models (dLLMs) as a multi‑granularity reconstruction space. Based on this insight, we propose FINDERS (FactorIzatioN of Diffusion statEs for Routing across Scenarios), a router‑equipped dLLM framework comprising four operations: denoise, factorize, route, and compress. Learnable meta‑slots organize complementary denoising states into transferable cross‑modal semantics and modality‑specific behavioral factors.
To provide grounded supervision, we construct BehaviorQA, a large‑scale corpus containing explicit behavioral rationales. Guided by scenario queries, a lightweight router dynamically emphasizes relevant slot directions via an input‑level residual pathway before cross‑source fusion and final compression.
Extensive offline experiments and online A/B tests across ten industrial scenarios show that FINDERS consistently outperforms state‑of‑the‑art methods in both unified and scenario‑adaptive settings. This work pushes dLLMs beyond language generation, demonstrating their immense potential as powerful user representation learners.
Review: FINDERS enables bidirectional cross‑modal information flow through diffusion denoising states and leverages a learnable routing mechanism for scenario‑aware representation compression, offering a more flexible and efficient approach to user modeling.