As large language models (LLMs) evolve rapidly, the demand for effective personalization grows. A personalized LLM must infer a user’s query‑dependent preferences from limited, heterogeneous, and continuously changing interaction histories. Existing solutions fall into three categories: in‑context personalization, which incurs recurring context‑management costs; parametric adaptation, which requires model optimization and deployment; and personalized reward modeling, which needs multi‑candidate generation and scoring. These approaches become especially restrictive when observations are sparse and preference evidence varies across queries and over time.
To tackle these challenges, we introduce \modelname, a fine‑grained, instance‑tailored steering framework. It dynamically creates sample‑level interference vectors from user data and injects them into the model’s forward pass for on‑the‑fly adaptation. The key components are:
- match contrast that extracts user‑relevant signals via contrastive learning;
- a fine‑grained steering component that hooks activations from attention and MLP layers to capture nuanced preferences;
- a query‑conditioned aggregation module that fuses these signals into contextually relevant enhancements.
Extensive experiments span short‑to‑long text generation and web function calling. Results consistently show significant performance gains, with extreme superiority in fast‑shifting and data‑sparse settings. Importantly, \modelname eliminates extra training cycles and adds only minimal storage and inference latency, making it practical for real‑world deployment. The implementation is publicly available.
Review