Intuition-Guided Latent Reasoning for LLM-Based Recommendation
Abstract Overview
This paper studies latent reasoning for LLM-based recommendation and argues that existing methods often begin reasoning from poorly aligned hidden states, which can lead to weaker recommendation trajectories. The proposed IntuRec framework addresses this by introducing recommendation intuition as a prior that anchors the initial latent reasoning state. IntuRec first extracts a top-K candidate set from an LLM recommender as an intuition source, then converts that set into a preference-aligned embedding through a dual-attention encoder and injects it into the reasoning start point. Experiments on three Amazon subsets (CDs, Toys, and Games) compare IntuRec against traditional recommenders and several LLM-based baselines, alongside ablation, visualization, and hyperparameter analyses.
Novelty
The distinctive idea is to treat recommendation intuition as an explicit latent prior for initializing LLM-based latent reasoning, rather than leaving the reasoning start point unconstrained. The method operationalizes this through a two-stage pipeline with Target-Aware Candidate Balancing for candidate construction and an Intuition Dual-Attention Encoder that transforms candidate items into a structured initialization embedding.
Results
Across the three evaluated datasets, IntuRec consistently improves over its underlying LLM recommendation backbones and generally delivers the strongest or near-strongest Recall and NDCG results among the compared methods. The ablation study shows that TACB, embedding-based intuition injection, self-attention, cross-attention, and the intuition alignment loss each contribute to performance, while visualization indicates that IntuRec's initialized reasoning states lie closer to target item embeddings than unconstrained latent reasoning baselines.
Key Points
- IntuRec introduces a two-stage intuition-guided latent reasoning framework for sequential recommendation, using top-K candidates as a structured prior for the reasoning start point.
- Empirical evaluation on CDs, Toys, and Games shows that the method improves over BIGRec, D3, and latent reasoning baselines, with ablations supporting the importance of each major component.
- Visualization and hyperparameter analysis suggest that the learned intuition embeddings provide a more target-aligned initialization, and that performance peaks with a moderate candidate set size and very few reasoning steps.