Large Language Models (LLMs) are increasingly recognized as the backbone for Generative Recommendation (GR), promising the utilization of pretrained world knowledge. However, reliably invoking this knowledge remains a challenge. A primary obstacle is that LLM-based GR typically uses Semantic IDs (SIDs) to represent items, which disrupts the natural language reasoning interface of LLMs due to these tokens being unseen during pretraining. Existing methods tackle this through costly multi-stage pipelines to ground SIDs and elicit explicit rationales, yet they offer limited insight into the necessity of each stage.
This work systematically decomposes explicit reasoning training pipelines for LLM-based GR, revealing three critical limitations: weakened verbalization of world knowledge, misalignment between SID and natural language token embedding spaces, and sensitivity to rationale quality, all hampering explicit reasoning performance. To address these issues, we propose PauseRec, a lightweight implicit reasoning paradigm tailored for GR.
PauseRec is exceptionally practical, avoiding costly reasoning trace acquisition and alignment training, leading to numerous benefits: (1) it outperforms standard explicit CoT methods by up to 6.22%, (2) reduces training cost by up to 65% GPU hours, and (3) speeds up inference by up to 71.3%. These results position PauseRec as a lightweight alternative to explicit rationale generation, enabling more effective and efficient LLM-based GR.
Blogger's Review: The introduction of PauseRec opens up new possibilities for the application of large language models in generative recommendation. Its significant performance gains and cost savings demonstrate the powerful potential of implicit reasoning. As technology continues to advance, future recommendation systems will be smarter and more efficient.