R3-REC: Reasoning-Driven Recommendation via Retrieval-Augmented LLMs over Multi-Granular Interest Signals

A research team led by Yuchen Miao has introduced R3-REC (Reasoning-Retrieval-Recommendation), a novel framework that tackles two persistent challenges in sequential recommendation systems. The system addresses evidence insufficiency caused by cold-start sparsity and noisy, length-varying item texts, while also providing transparency into dynamic, multi-faceted user intents across different time horizons. By unifying five key components—Multi-level User Intent Reasoning, Item Semantic Extraction, Long-Short Interest Polarity Mining, Similar User Collaborative Enhancement, and Reasoning-based Interest Matching and Scoring—R3-REC creates a comprehensive approach to understanding user preferences.

In rigorous testing across ML-1M, Games, and Bundle datasets, R3-REC consistently outperformed strong neural and LLM baselines, delivering improvements of up to +10.2% in HR@1 (hit rate at top-1 recommendation) and +6.4% in HR@5 metrics. The framework maintains manageable end-to-end latency despite its sophisticated reasoning capabilities, making it practical for real-world deployment. Ablation studies confirmed that all five modules contribute complementary gains to the system's overall performance, with the paper accepted for presentation at the prestigious ICASSP 2026 conference.

The framework's prompt-centric, retrieval-augmented architecture allows it to overcome traditional limitations where sparse user data or noisy item descriptions would degrade recommendation quality. By analyzing interest signals at multiple granularities and incorporating collaborative filtering through similar user enhancement, R3-REC provides more accurate and explainable recommendations than previous approaches. This represents a significant advancement in making large language models effective for personalized recommendation tasks beyond their traditional text generation capabilities.