Cross-domain sequential recommendation (CDSR) tries to ease data sparsity by transferring user behavior patterns from one domain to another. Recent work has increasingly turned to large language models (LLMs) to create a bridge between domains, using global item semantics as a shared foundation.
The new arXiv paper, "When Semantics Fall Short," pushes back on that reliance on semantic bridging. Its title signals a core argument: item-level semantics alone are insufficient for effective transfer. Instead, the paper proposes a persona-guided method, suggesting that modeling stable user traits or preferences is key to making cross-domain recommendations work.
The abstract is truncated in the source, so full methodological details and experimental results are not available. What is clear from the framing is that the paper positions itself against purely semantic LLM-based transfer and emphasizes the role of user personas in overcoming the limits of semantics.