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..

MURKA: Multi-Reward Reinforcement Learning with Knowledge Alignment for Optimization Tasks

Authors: WANTONG XIE, Yi-Xiang Hu, Jieyang Xu, Feng Wu, Xiangyang Li

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate MURKA s generalizability through extensive experiments across diverse OR benchmarks, demonstrating its robustness and scalability. Experimental results on eight diverse OR benchmarks, including NLP4LP, Complex OR, and NL4Opt, demonstrate that MURKA, built on the LLa Ma3-8B backbone, achieves a 5.9% absolute improvement in solution accuracy and a 5.1% increase in execution success rate compared to leading baselines.
Researcher Affiliation Academia 1Institute of Advanced Technology, University of Science and Technology of China 2School of Computer Science and Technology, University of Science and Technology of China EMAIL EMAIL
Pseudocode Yes Algorithm 1 Comprehensive Reward Calculation Algorithm 2 Structure Validation Reward Algorithm 3 Content Validation Reward
Open Source Code Yes The supplemental material includes the code, a README file, training example data, and log examples, providing sufficient instructions to reproduce the main experimental results as described in Appendix D.
Open Datasets Yes We evaluated our method on eight diverse benchmarks NL4Opt [Ramamonjison et al., 2023], Mamo Easy, Mamo Complex [Huang et al., 2024], NLP4LP [Ahmadi Teshnizi et al., 2024], Complex OR [Xiao et al., 2023], Industry OR [Huang et al., 2025], Opti Bench [Yang et al., 2024], and Opt MATH [Lu et al., 2025] comprising 2,224 problem instances.
Dataset Splits Yes From the test set, we randomly sampled 20% of the instances, stratified by scenarios and types, to serve as the foundation for knowledge distillation and data synthesis. ... ensuring complete separation between training and test sets to prevent data leakage.
Hardware Specification Yes Appendix D.2 specifies the use of eight NVIDIA Ge Force RTX 4090 GPUs with 24 GB memory each and four AMD EPYC 7763 64-Core Processors @ 2.45GHz.
Software Dependencies Yes To solve the optimization model, we use Gurobi 12.0.1 [Gurobi Optimization, LLC, 2025].
Experiment Setup Yes Table 12: Hyperparameter Configuration for Extractor and Solver. Model Epoch Batch Learning Rate Lo RA_R Max_Length Warm Up Ratio Weight Decay Adam Beta E 25 16 5e-6 16 1024 0.1 0.1 0.9 S 15 32 5e-5 8 3072 0 0 0.9