Escaping the Euclidean Void: Manifold-Informed Flow Matching for Sequential Recommendation

Sequential recommendation models often use generative flow matching to synthesize target item representations. However, these models frequently suffer from the Euclidean void—a phenomenon where the straight-line path between source noise and a target item traverses regions of the continuous embedding space that lack valid item semantics. This paper asks: how can we ensure that generative trajectories remain semantically grounded without sacrificing the computational efficiency of one-step inference?

The authors propose MIRAGE (Manifold-Informed Rectification framework for Accelerated Generation of Embeddings). Instead of altering the straight probability path, MIRAGE reorganizes the underlying embedding geometry during training. It uses an item co-occurrence graph as a proxy for the semantic manifold. By applying a time-modulated topology regularizer, the model aligns interpolated path states with local graph anchors. This process dynamically repositions item embeddings to form a cohesive manifold around the generative trajectory. Crucially, this rectification is performed only during training; the inference phase remains a standard, efficient one-step process that does not require graph lookups.

Experimental results across four real-world datasets demonstrate that MIRAGE consistently outperforms state-of-the-art generative recommendation baselines. The model is particularly effective at improving performance for long-tail items, where direct supervision is sparse. By ensuring that the generative path is semantically supported, MIRAGE achieves higher overall accuracy while maintaining the efficiency benefits of one-step transport.

This work shifts the focus in generative recommendation from merely optimizing target scores to ensuring the reliability of the entire generative process. By identifying the Euclidean void as a structural failure, the authors provide a novel way to bridge the gap between continuous latent space models and discrete item catalogs. This approach offers a robust way to improve recommendation quality without increasing the computational burden during deployment.