Personalizing Large Language Models (LLMs) for long-term multi-turn interactions remains challenging, as user preferences and behavioral traits are progressively revealed and continuously evolving throughout conversations. Previous personalization methods have mostly relied on prompt-based or post-training (e.g., fine-tuning) paradigms; however, these approaches have been either prone to shallow alignment or have suffered from high computational costs. Inspired by the recent emergence of activation steering, we explore whether personalized alignment can be achieved through lightweight latent-space intervention without modifying backbone parameters. To achieve this goal, we propose PersonaSteer, a dynamic side-model architecture for multi-turn personalized alignment via learnable activation steering. Specifically, to enable fine-grained and evolving personalization, our design learns high-dimensional steering representations from unstructured interaction history and injects them into the backbone model’s residual stream for latent-space behavioral modulation. Furthermore, we introduce a steering-vector feedback loop to maintain persona consistency across multi-turn conversations. Extensive experiments demonstrate that PersonaSteer achieves best or top-tier alignment performance, averaging 3.18/5 on Qwen2.5-7B and 2.95/5 on Qwen2.5-3B. In addition, our framework approach maintains stability during evolving dialogues, with alignment performance consistently improving over time. PersonaSteer provides a lightweight solution for building truly dynamic and personalized LLM. The code is available in https://github.com/tsinghua-fib-lab/PersonaSteer.

Large language models (LLMs) have demonstrated strong capabilities in complex reasoning tasks, yet their autoregressive generation paradigm makes it computationally expensive to explore diverse reasoning paths. In contrast, diffusion language models (DLMs) adopt a parallel, non-autoregressive generation mechanism that enables the efficient production of diverse candidate outputs. Motivated by this complementarity, we explore a collaborative reasoning framework that combines diffusion-based generation with autoregressive evaluation. Specifically, we leverage DLMs to efficiently generate diverse intermediate reasoning thoughts, and employ LLMs as evaluators to assess and select candidates based on their plausibility and correctness. By decoupling proposal generation from evaluation, our framework exploits the strengths of both models: efficient exploration from diffusion models and causally grounded assessment from autoregressive models, which naturally aligns with the divergent-convergent thinking framework in cognitive psychology. Experiments across various mathematical and logical reasoning benchmarks demonstrate that our framework improves inference efficiency while maintaining competitive or superior reasoning accuracy, laying the groundwork for building efficient reasoning architectures. Our code is open-source at https://anonymous.4open.science/r/Diffuse-Thinking-EC60.