Multimodal large language models (MLLMs) have achieved remarkable progress in multimodal reasoning, yet they still suffer from hallucinations that undermine their reliability. Existing inference-time mitigation methods mainly focus on optimizing decoding strategies or analyzing internal attention mechanisms, while the role of visual representations remains underexplored. In this work, we reveal that visual representations exhibit information dispersion during layer-wise evolution, manifested as a progressive deviation from the initial visual subspace and increasing alignment with the textual subspace. Based on this insight, we propose Visual Representation Regularization (VRR), a training-free method that mitigates hallucinations by regularizing visual representations toward their initial states. Extensive experiments demonstrate that VRR consistently reduces hallucinations and is compatible with existing mitigation methods. Our code is available at https://github.com/Jiarker/VRR.