IntuRec uses cognitive intuition to guide LLM latent reasoning for recommendations

Large Language Models (LLMs) show promise in recommendation, but existing latent reasoning methods start from unconstrained hidden representations misaligned with target items. Inspired by cognitive neuroscience—where human reasoning uses intuitive priors—researchers from multiple institutions developed IntuRec. It introduces a two-stage process: first, an extraction stage generates a top-K candidate set from user history as 'recommendation intuition'. Second, an injection stage transforms that set into a preference-aligned intuition embedding using self- and cross-attention mechanisms. This embedding initializes the reasoning start point, guiding latent reasoning along more accurate trajectories in continuous hidden space.

Experiments on multiple real-world datasets show IntuRec consistently outperforms state-of-the-art baselines, confirming that grounding LLM reasoning with semantic priors improves recommendation quality. The code is open-sourced. By mimicking the brain's intuitive leap before deliberate reasoning, IntuRec offers a practical blueprint for making LLM-based recommenders both faster and more accurate—without requiring additional training data or massive compute.