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

Explain Like I'm Normal

Researchers are ditching standard LLM user profiles for 'FINDERS,' a system that uses diffusion language models to encode user behavior. Instead of one-size-fits-all summaries, it treats user data as an iterative denoising process, allowing the model to extract different levels of detail depending on the specific task. This solves the 'static representation' problem where information is lost during compression.