FleetPilot: Fine-Grained Steering Empowers Real-World LLM Personalization

Key points

• With the rapid evolution of large language models (LLMs), there is an increasing demand for effective personalization techniques that can adapt model behavior to individual user preferences.1
• Personalized LLMs must infer a user's query-dependent preferences from limited, heterogeneous, and evolving interaction histories.1
• Our approach introduces: match contrast that produces user-relevant signals, fine-grained steering component that captures nuanced signals by hooking activations from attention and MLP layers, and query-conditioned aggregation module that synthesizes these signals into contextually relevant enhancements.1
• Additionally, \modelname rids us of training cycles and incurs light overhead of storage or inference latency.1

Sources (1)

• [1]FleetPilot: Fine-Grained Steering Empowers Real-World LLM Personalization
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 02:44 PM

  To address these challenges, we propose \modelname, a fine-grained and instance-tailored steering framework that dynamically generates sample-level interference vectors from user data and injects them into the model's forward pass for personalized adaptation.

  With the rapid evolution of large language models (LLMs), there is an increasing demand for effective personalization techniques that can adapt model behavior to individual user preferences.

Extractive summary: sentences quoted from the sources.