SteerX: Disentangled Steering for LLM Personalization

TL;DR

SteerX tackles the problem that activation steering for LLM personalization often overfits to noisy, non-preference content in full user histories. It introduces a three-stage pipeline—disentangling tokens via token-level causal effects, smoothing them into coherent style descriptions, and steering with these descriptions—grounded in causal inference. Empirical results show SteerX consistently improves steering vector quality and personalization across two backbones and real-world datasets, outperforming both non-personalized baselines and existing steering methods. The approach offers a practical, interpretable, and broadly applicable framework for more accurate and robust LLM personalization, with potential extensions to multimodal and real-time contexts.

Abstract

Large language models (LLMs) have shown remarkable success in recent years, enabling a wide range of applications, including intelligent assistants that support users' daily life and work. A critical factor in building such assistants is personalizing LLMs, as user preferences and needs vary widely. Activation steering, which directly leverages directions representing user preference in the LLM activation space to adjust its behavior, offers a cost-effective way to align the model's outputs with individual users. However, existing methods rely on all historical data to compute the steering vector, ignoring that not all content reflects true user preferences, which undermines the personalization signal. To address this, we propose SteerX, a disentangled steering method that isolates preference-driven components from preference-agnostic components. Grounded in causal inference theory, SteerX estimates token-level causal effects to identify preference-driven tokens, transforms these discrete signals into a coherent description, and then leverages them to steer personalized LLM generation. By focusing on the truly preference-driven information, SteerX produces more accurate activation steering vectors and enhances personalization. Experiments on two representative steering backbone methods across real-world datasets demonstrate that SteerX consistently enhances steering vector quality, offering a practical solution for more effective LLM personalization.

Figures (8)

• Figure 1: Two representative training-free activation steering methods. Influence Vector is the logit difference with vs. without history. Style Vector is the mean hidden-state difference between user-authentic response $h_i$ and user-agnostic response $\bar{h}_i$.
• Figure 2: Illustration of our causal inference framework for personalized text generation.
• Figure 3: Overview of the proposed SteerX framework for personalized text generation. (Left) Personalized anchor token identification: For each history $h_i$, we apply a leave-one-history-out strategy and mark tokens with causal influence scores exceeding a predefined threshold as personalized tokens. (Middle) Style profile generation: All personalized tokens are transformed from a discrete token set into a continuous, coherent preference description$S$. (Right) Personalization steering: $S$ is injected into the generation process. A personalization scaling factor $\lambda$ controls the strength of steering based on the style vector.
• Figure 4: Ablation results on the topic writing task, comparing the full SteerX framework with two reduced variants: (1) w/o Smoothing and (2) w/o Personalized Token. Performance is reported across ROUGE-1, METEOR, BLEU, and BERTScore.
• Figure 5: Effects of different hyperparameters on SteerX (COS) for the topic writing task: (1) upper panel — varying the causal effect threshold $\tau$; (2) lower panel — varying the steering coefficient $\lambda$. Performance is evaluated using four metrics: ROUGE-1, METEOR, BLEU, and BERTScore.