Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs

Authors: Xiaqiang Tang, Jian Li, Nan Du, Sihong Xie

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on two benchmark KGQA datasets demonstrate that our method significantly outperforms baseline methods in non-stationary settings while achieving state-of-the-art performance in stationary environments.
Researcher Affiliation Collaboration 1The Hong Kong University of Science and Technology (Guangzhou) 2Tencent Hunyuan
Pseudocode Yes Algorithm 1: Deep GGI-MO bandit enhanced RAG learning algorithm
Open Source Code Yes Code https://github.com/FUTUREEEEEE/Dynamic-RAG
Open Datasets Yes We evaluate our systems on two KGQA datasets Web QSP (Yih et al. 2016) and Complex Web Questions (CWQ) (Talmor and Berant 2018)
Dataset Splits No The paper does not explicitly provide train/test/validation splits within the main text for each dataset. While it mentions training on Web QSP and testing on CWQ for one scenario, it doesn't specify the splits *within* Web QSP or CWQ datasets for general experiments.
Hardware Specification Yes All experiments are conducted on the Nvidia Tesla V100 graphical card with the Intel Xeon Platinum 8255C CPU.
Software Dependencies No The paper mentions software components like Llama-2-7b-chat-hf and Llama Index but does not provide specific version numbers for them (e.g., 'Llama Index 2024' refers to the documentation year, not a software version).
Experiment Setup No The main text refers to an appendix for detailed setup ('See subsection 3 in the appendix (Tang et al. 2024) for detail set up.') but does not provide specific hyperparameters like learning rate, batch size, or optimizer settings in the main content.