@article{Zhang2026, 
author = {Zixuan Zhang and Bowen Hao and Yurui Wang and Xiang Wei and Qimeng Niu and Yuxuan Li},
title = {FT-HashRAG: Combining Hash Retrieval-Augmented Generation with Fine-Tuning for Universal Explainable Recommendation},
year = {2026},
journal = {Tsinghua Science and Technology},
volume = {31},
number = {4},
pages = {2275-2291},
keywords = {explainable recommendation, fine-tuning, Large Language Model (LLM), Hash Retrieval-Augmented Generation (HashRAG)},
url = {https://www.sciopen.com/article/10.26599/TST.2025.9010106},
doi = {10.26599/TST.2025.9010106},
abstract = {Explainable recommendation have garnered significant research interest for their transparency. Recent advancements employ Large Language Models (LLMs) through two primary approaches: LLM-auxiliary models, which improve accuracy but rely on conventional recommendation models and produce limited explanations, and LLM-based models, which employ fine-tuning or Graph Retrieval-Augmented Generation (GraphRAG) for explainability——yet face limitations in domain adaptability and neglect intrinsic signals like item weights. Moreover, existing methods struggle with multi-scenario generalization. To address these challenges, we propose FT-HashRAG, a universal explainable recommendation framework that integrates Hash Retrieval-Augmented Generation (HashRAG) with fine-tuning. HashRAG constructs user-specific blocks using intrinsic (e.g., item weights) and extrinsic signals (e.g., item descriptions) to generate fine-grained explanations. By integrating HashRAG with fine-tuning, FT-HashRAG enhances the LLM’s capacity to synthesize retrieved information effectively. We curate a multi-scenario instruction dataset including four explainable recommendation scenarios: item weights, item attributes, popularity bias, and interaction paths, and train LLM via Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Experiments show FT-HashRAG significantly outperforms state-of-the-art baselines. We open-source our model on https://huggingface.co/tczzx6/FT-HashRAG.}
}