Latent Logic Tree Extraction for Event Sequence Explanation from LLMs

Authors: Zitao Song, Chao Yang, Chaojie Wang, Bo An, Shuang Li

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical demonstrations showcase the promising performance and adaptability of our framework. Empirical results show that this method notably enhances generalization in event histories with semantic information.
Researcher Affiliation Collaboration 1School of Computer Science and Engineering, Nanyang Technological University, Singapore 2School of Data Science, The Chinese University of Hong Kong, Shenzhen, China 3Skywork AI, Singapore.
Pseudocode Yes Algorithm 1 Bayesian Logic Tree Learning for Events
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes MIMIC-III (Johnson et al., 2016), an electronic health record dataset from intensive care unit patients. ... EPIC-KITCHENS-100 (EPIC-100) (Damen et al., 2021), which documents everyday kitchen activities... Stack Overflow (SO) (Leskovec & Krevl, 2014), which records a sequence of reward history...
Dataset Splits Yes We consider each sequence as a record pertaining to a single individual and partition each dataset into 80%, 10%, 10% train/dev/test splits by the total population.
Hardware Specification Yes All the experiments were conducted on a server with 512G RAM, two 64 logical cores CPUS (AMD Ryzen Threadripper PRO 5995WX 64-Cores), and four NVIDIA RTX A6000 GPUs with 50G memory.
Software Dependencies No The paper states 'All models are implemented using the Py Torch framework.' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We present the selected hyperparameters on synthetic datasets and three real-world datasets in Table 6 and Table 7 respectively. These tables list specific values for 'EPOCHS', 'BATCH SIZE', 'LLM LR', 'LOGIC TREE DEPTH', 'LOGIC TREE WIDTH', and other training parameters.