Research map / Learning theory and algorithms / Combinatorial and search optimization
FMIP: Joint Continuous-Integer Flow For Mixed-Integer Linear Programming
Hongpei Li, Hui Yuan, Han Zhang, Jianghao Lin, Dongdong Ge, Mengdi Wang, Yinyu Ye
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model joint distribution of integer and continuous variables (add continuous generation) (ablation) | FMIP · Item Placement, ND 400s | primal gap: 0.69 -> 0.25 | Improved |
| add feasibility guidance (ablation) | FMIP · Item Placement, ND 400s | primal gap: 1.01 -> 0.25 | Improved |
| add objective guidance (ablation) | FMIP · Item Placement, PMVB 600s | primal gap: 8.58 -> 0.67 | Improved |
| add holistic guidance (both terms) at inference (ablation) | FMIP · Item Placement, ND 400s | primal gap: 0.84 -> 0.25 | Improved |
| add holistic guidance at inference (ablation) | FMIP · Item Placement, PS 600s | primal gap: 0.24 -> 0.18 | Improved |
| use FMIP instead of best learning baseline | downstream solvers (ND, PS, PMVB, Apollo) · eight MILP benchmarks | primal gap: 41.34% average relative improvement | Improved |
| use FMIP on simple benchmarks where baselines already reach optimum (CA, GIS, MIS, FCMNF with PS/PMVB/Apollo) | downstream solvers · CA/GIS/MIS/FCMNF | primal gap: 0.00% relative improvement in many cells | No clear effect |
| swap graph backbone (Bi-GCN, Tri-GCN, GAT, ClusterGCN) under FMIP (ablation) | FMIP · IP and LB, objective value | objective value: IP ND: 14.11, 13.99, 13.95, 14.06 | No clear effect |
| increase number of sampling steps beyond an optimum (ablation) | FMIP · downstream solvers | primal gap | Hurt |
| add adaptive guidance balancing at 32 steps (ablation) | FMIP · Item Placement, PS | objective value: 13.25 -> 11.92 | Improved |
| add adaptive guidance balancing at 12 steps (ablation) | FMIP · Item Placement, PS | objective value: 11.79 -> 11.93 (alpha=1.1) | No clear effect |
| apply FMIP to very large MIPLIB instances (70,000 to 300,000 variables) | FMIP · MIPLIB 2017 size subsets | relative primal gap: 25.38% -> 23.26%; gap stays large | Mixed |
Later papers that compared with this method
1 later accepted paper reports FMIP as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Unsupervised Neural Langevin Sampler for Mixed Integer Linear Programming (ICML 2026 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Towards Foundation Models for Mixed Integer Linear Programming (ICLR 2025 · similarity 0.90)
- Apollo-MILP: An Alternating Prediction-Correction Neural Solving Framework for Mixed-Integer Linear Programming (ICLR 2025 · similarity 0.89)
- Fast and Interpretable Mixed-Integer Linear Program Solving by Learning Model Reduction (AAAI 2025 · similarity 0.89)
- MILP-StuDio: MILP Instance Generation via Block Structure Decomposition (NeurIPS 2024 · similarity 0.87)
- Contrastive Predict-and-Search for Mixed Integer Linear Programs (ICML 2024 · similarity 0.87)
- Generalization Guarantees for Learning Score-Based Branch-and-Cut Policies in Integer Programming (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
