RecSys paper introduces structure-preserving projection to fix modality bias in LLM recommenders

LLM-based recommender systems have become popular by projecting collaborative embeddings—vectors that capture user-item interaction patterns—into the embedding space of large language models. This lets them combine textual understanding with collaborative signals. However, that projection step introduces modality bias: the transformation can distort the relational geometry of the original collaborative embeddings, making the projected embeddings less useful for recommendation. The result is a mismatch between what the LLM understands and the true collaborative structure.

To solve this, Chiu and colleagues introduce a structure-preserving projection approach. Instead of allowing an unconstrained mapping, they apply structure-preserving losses that explicitly maintain the pairwise distances and neighborhood relationships of collaborative embeddings during projection. This keeps the geometry intact while still aligning with the LLM's semantic space. In comprehensive experiments across recommendation datasets, the method consistently outperforms baseline LLM-based recommenders. Accepted at RecSys 2026, this work provides a practical fix for a core limitation in current hybrid recommenders, making LLM-based sequential recommendation more accurate and trustworthy.