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..
Learning to Condition: A Neural Heuristic for Scalable MPE Inference
Authors: Brij Malhotra, Shivvrat Arya, Tahrima Rahman, Vibhav G Gogate
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our learned heuristic significantly reduces the search space while maintaining or improving solution quality over state-of-the-art methods. |
| Researcher Affiliation | Academia | Brij Malhotra Department of Computer Science The University of Texas at Dallas EMAIL Shivvrat Arya Department of Computer Science New Jersey Institute of Technology EMAIL Tahrima Rahman Department of Computer Science The University of Texas at Dallas EMAIL Vibhav Gogate Department of Computer Science The University of Texas at Dallas EMAIL |
| Pseudocode | Yes | Algorithm 1: Data Collection for L2C with Conditioning Assignments Algorithm 2: Beam Search for L2C |
| Open Source Code | Yes | Our implementation, solver integrations, and experiment scripts are publicly available at, https://github.com/brijml/L2C. |
| Open Datasets | Yes | We evaluated our method and baselines on 14 high-treewidth binary probabilistic graphical models (PGMs) from UAI inference competitions [44, 45] |
| Dataset Splits | Yes | Using Gibbs sampling [1], we generated 12,000 training, 1,000 test, and 1,000 validation examples per model. |
| Hardware Specification | Yes | All models were implemented in Py Torch [47] and executed on an NVIDIA A40 GPU. |
| Software Dependencies | Yes | We use two oracles on unconditioned queries and as final solvers post-conditioning: the SCIP solver [23] ... [23] Suresh Bolusani et al. The SCIP Optimization Suite 9.0. Technical report, Optimization Online, 2024. |
| Experiment Setup | Yes | The neural networks in L2C-OPT and L2C-RANK use 256-dimensional embeddings, two multi-head attention layers, and 15 skip-connection blocks. Dense layers have 512 units with 0.1 dropout [41] and Re LU activations. We trained the models using Adam [46] with a learning rate of 8 10 4, an exponential decay rate of 0.97, a batch size of 128, and early stopping after 5 stagnant epochs (maximum 50 epochs). |