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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
RoME: Domain-Robust Mixture-of-Experts for MILP Solution Prediction across Domains
Authors: Tianle Pu, Zijie Geng, Haoyang Liu, Shixuan Liu, Jie Wang, Li Zeng, Chao Chen, Changjun Fan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct comprehensive experiments to evaluate the effectiveness of Ro ME in learning a unified model for solving MILP problems from several different domains, including those in-distribution and out-of-distribution datasets, and zero-shot generalization to real-world instances from MIPLIB. We then conduct experiments to analyze the contributions of different components, as well as the featured patterns of our learned model. |
| Researcher Affiliation | Academia | 1Laboratory for Big Data and Decision, College of Systems Engineering, National University of Defense Technology 2Mo E Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China 3College of Computer Science and Technology, National University of Defense Technology EMAIL EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using textual explanations and mathematical equations in Section 3, but does not present any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/happypu326/Ro ME. |
| Open Datasets | Yes | We conduct experiments on five MILP problem families widely studied in the literature: Independent Set (IS), Set Covering (SC), Item Placement (IP), Combinatorial Auctions (CA), and Workload Appointment (WA). ... The IP and WA datasets are from Neur IPS 2021 ML4CO competition [49]. ... We evaluate it on a curated subset of the MIPLIB benchmark [10]. |
| Dataset Splits | Yes | To train this model, we use a training dataset composed of 720 instances, where the dataset contains 240 IS instances, 240 IP instances and 240 SC instances. The validation set contains 240 instances with 80 IS, IP and SC instances, respectively. ... For evaluation, we run all the methods on 100 testing instances for each dataset. |
| Hardware Specification | Yes | All experiments are conducted on a single machine equipped with NVIDIA Ge Force RTX 3090 GPUs and AMD EPYC 7402 24-core CPUs running at 2.80GHz. |
| Software Dependencies | Yes | We use Gurobi version 11.0.3 and SCIP version 8.1.0 in all experiments. |
| Experiment Setup | Yes | During training, we set the initial learning rate to 0.0005 and train the model for 10,000 epochs, employing an early stopping mechanism to prevent overfitting. In the testing phase, the parameters (k0, k1, θ) determine which variables are fixed to 0 or 1 and define the scope of the subsequent search process. These parameters used in this work are detailed in Table 3. ... The key training parameters are summarized in Table 4. |