Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Wait-Less Offline Tuning and Re-solving for Online Decision Making

Authors: Jingruo Sun, Wenzhi Gao, Ellen Vitercik, Yinyu Ye

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate at least 10-fold improvements in regret over firstorder methods and 100-fold improvements in runtime over LP-based methods. ... We conduct extensive experiments to evaluate our algorithm s performance and validate our theoretical results. This section is divided into two parts. In the first part (Section 4.1), we evaluate Algorithms 1 and 2 across different choices of re-solving frequency. In the second part (Section 4.2), we compare our algorithm with LP-based and first-order methods in terms of regret and running time.
Researcher Affiliation Academia 1 Department of Management Science & Engineering, Stanford University 2 Institute for Computational and Mathematical Engineering, Stanford University 3 Department of Computer Science, Stanford University 4 Chinese University of Hong Kong (Shen Zhen) 5 Hong Kong University of Science and Technology. Correspondence to: Jingruo Sun <EMAIL>.
Pseudocode Yes Algorithm 1 Parallel Multi-Phase OLP Algorithm ... Algorithm 2 Enhanced Multi-Start OLP Algorithm
Open Source Code Yes All implementations can be found at Git Hub Link.
Open Datasets No We consider the following distributions: Input I: ait Unif[0, 2], rt Unif[0, 10] Input II: ait N(0.5, 1), rt N(0.5m, m) ... Input III: ait min(1, max(0, 1 + z)), rt Unif[0, 1] where z t(1) : Student s t-distribution with 1 degree of freedom. ... We generate {rt, at}T t=1 from Input III
Dataset Splits No The paper describes generating sequences of data points {(rt, at)}T t=1 randomly from uniform and normal distributions over a time horizon T, but does not specify any explicit train/test/validation splits for the generated data.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only discusses runtime performance generally.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1).
Experiment Setup Yes Learning rates are selected as specified in Section 3: Algorithm 1: αt = (O(1/f 1/2) t f O(1/f 2/3) t kf Algorithm 2: αt = O(1/t) ... We set the resource types to m = 5, time T to range evenly over [102, 106], average resource di Uniform[1/3, 2/3], and re-solving frequency to f = T 1/3 from Section 4.1.