WhisperRec: Latent Reasoning for Efficient Foundation Recommendation Models

A recent study presents WhisperRec, a novel framework designed for efficient latent reasoning in foundation recommendation models (FRMs). While large language models (LLMs) have been utilized as foundational elements for FRMs due to their reasoning abilities, current methods employing explicit Chain-of-Thought (CoT) within the Think-then-Answer framework face significant inference delays and inflexible templates. WhisperRec transforms teacher-generated CoT into learnable latent reasoning tokens, facilitating a Latent-Reason-then-Answer approach that conducts reasoning in latent space without lengthy explanations. This innovation mitigates latency issues while preserving critical decision-making information. Additionally, the framework features Multi-View Adaptive CoT to capture a variety of user preferences. The paper can be accessed on arXiv under ID 2607.26621.