ICLR 2026

Joint Continuous-Integer
Flow for MILP

The first generative framework that models the joint distribution of both integer and continuous variables for Mixed-Integer Linear Programming solutions.

41.34% Average Improvement
8 Benchmarks
SOTA Performance
Hongpei Li1 Hui Yuan2 Han Zhang3 Jianghao Lin4* Dongdong Ge4 Mengdi Wang2 Yinyu Ye5
1Shanghai University of Finance and Economics · 2Princeton University · 3National University of Singapore · 4Shanghai Jiao Tong University · 5Stanford University
FMIP Framework

Abstract

Mixed-Integer Linear Programming (MILP) is a foundational tool for complex decision-making problems. However, the NP-hard nature of MILP presents significant computational challenges.

While recent generative models have shown promise, they suffer from a critical limitation: they model the distribution of only the integer variables and fail to capture the intricate coupling between integer and continuous variables, creating an information bottleneck and ultimately leading to suboptimal solutions.

We propose FMIP, the first generative framework that models the joint distribution of both integer and continuous variables for MILP solutions. Built upon the joint modeling paradigm, a holistic guidance mechanism steers the generative trajectory toward optimality and feasibility during inference.

Key Achievement

41.34%

Average relative improvement over existing baselines across 8 standard MILP benchmarks.

Why FMIP is Different

Compared to existing methods that only model integer variables

Existing Methods
Integer Only
Partial Solution
  • Model only integer variables
  • Ignore continuous variables
  • Information bottleneck
  • Limited guidance capability
VS
FMIP (Ours)
Integer + Continuous
Complete Solution
  • Joint distribution modeling
  • Full variable coupling
  • Holistic guidance
  • Superior performance
FMIP vs Existing Methods

FMIP jointly models both variable types with holistic guidance, while existing methods only handle integer variables.

What Makes FMIP Special

Joint Distribution

First framework to model complete distribution of both integer and continuous variables

01

Holistic Guidance

Steers sampling toward optimality and feasibility using complete solution feedback

02

Universal Compatible

Works with arbitrary backbone networks and various downstream solvers

03

SOTA Performance

Superior results on 8 benchmarks with 41.34% average improvement

04

Experimental Performance

8 Benchmark Datasets
4 Downstream Solvers
4+ Backbone Networks
41% Avg. Improvement

Relative Improvement over Best Baseline

FMIP achieves 41.34% average improvement across 8 benchmarks and 4 downstream solvers

Dataset ND
(400s)
PS
(600s)
PMVB
(600s)
Apollo
(800s)
CA Auctions 50.6% 0.0% 0.0% 0.0%
GIS Independent Set 4.1% 14.3% 0.0% 0.0%
MIS Max Independent Set 95.9% 0.0% 100% 100%
FCMNF Network Flow 3.5% 0.0% 0.0% 0.0%
SC Set Covering 60.0% 100% 97.9% 100%
LB Load Balancing 93.8% 100% 0.0% 100%
IP Item Placement 71.6% 74.9% 48.1% 100%
MIPLIB Standard 0.4% 0.5% 0.6% 0.3%
Average 47.2% 23.7% 30.8% 37.5%
>50% improvement
10-50% improvement
0-10% improvement
Tied for best

Datasets Covered

Combinatorial Auctions Generalized Independent Set Maximum Independent Set Fixed-Charge Multi-Commodity Flow Set Covering Load Balancing Item Placement MIPLIB 2017

Open Source & Ready to Use

FMIP is freely available on GitHub. Try it on your MILP problems today!

View on GitHub

BibTeX

@misc{li2025fmipjointcontinuousintegerflow,
      title={FMIP: Joint Continuous-Integer Flow For Mixed-Integer Linear Programming}, 
      author={Hongpei Li and Hui Yuan and Han Zhang and Jianghao Lin and Dongdong Ge and Mengdi Wang and Yinyu Ye},
      year={2025},
      eprint={2507.23390},
      archivePrefix={arXiv},
      primaryClass={math.OC},
      url={https://arxiv.org/abs/2507.23390}, 
}