Latent-Aligned Reasoning for Multimodal Recommendation

To address the issue of audio-visual signal attenuation (cross-modal dilution) in recommendation tasks due to multi-step reasoning in multimodal visual language models, researchers proposed a latent alignment reasoning framework called LARK. This framework consists of two stages: In the first stage, learnable implicit tokens are intertwined with multi-step reasoning chains and aligned with a frozen visual encoder to preserve perceptual details; in the second stage, full-connected layer projections are used for training, leveraging contrastive learning between items to prevent semantic fade, while aligning intermediate features with the hidden states of the reasoning chain from the first stage. Experiments show that LARK achieves the most advanced performance across various recommendation architectures on three public benchmark datasets and one industrial dataset, and ablation experiments confirm the independent contribution of each component.