Sequential recommendation is crucial for predicting user intent from historical behaviors. However, existing methods are constrained by static representations that compress a user’s history into a single vector, lacking the reasoning capability to capture dynamic user intents. To address this limitation, we introduce CARARec, a unified framework that embodies a Dynamic Reasoning paradigm. This paradigm models user intent as a sequential trajectory of “thoughts”, where each step progressively refines the user representation within a continuous latent hyper action space. The framework is built on three core components to solve key challenges: (1) Latent Concept Synthesis (LCS) anchors the reasoning process within the item vocabulary, ensuring semantic coherence; (2) Coarse-to-Fine Concept Selection (CFCS) makes this process computationally tractable; and (3) Reasoning Path Alignment (RPA) imposes a coherent and generalizable logic. Extensive experiments on real-world benchmarks demonstrate that CARARec substantially outperforms state-of-the-art baselines. Moreover, ablation studies confirm the necessity of each component, highlighting the principled and effective design of the framework. Notably, the benefits of CARARec are architecture-agnostic, enabling seamless integration with various backbone models. The specific implementation of CARARec is available in the code repository: https://github.com/stephen863256/CARARec .