NextQuill: Causal Preference Modeling for Enhancing LLM Personalization

A research team led by Xiaoyan Zhao, Juntao You, and others has published NextQuill, a new framework for personalizing large language models (LLMs) like GPT-4 and Claude using causal inference. The core innovation is treating both model predictions and training data generation as outcomes influenced by user preferences alongside other confounding factors. NextQuill employs causal intervention techniques to estimate the "true preference effect"—the specific impact of a user's history on each token prediction—rather than superficially aligning model outputs with all available data.

NextQuill implements two complementary alignment strategies. First, it aligns the model's internal causal preference effects on its predictions with those reflected in ground-truth data. Second, it focuses training on fitting only the "preference-bearing" tokens identified through this causal analysis, ignoring tokens that don't genuinely reflect user intent. This method shifts personalization from brute-force pattern matching to learning the underlying causal structure of user preferences.

Experiments across multiple personalization benchmarks demonstrate that NextQuill significantly outperforms existing methods in personalization quality. The framework, accepted for ICLR 2026, provides a more robust, interpretable foundation for adapting LLMs to individual users in applications like personalized assistants, content recommendation, and adaptive learning systems. The code is publicly available, enabling integration into existing model fine-tuning pipelines.