MILP-Evo: LLM сам проектирует и ускоряет логику MILP-солверов оптимизации

Researchers have presented MILP-Evo — a system in which a large language model (LLM), working in a closed loop, writes, runs, and improves the logic of mixed-integer linear programming (MILP) solvers on its own. The preprint appeared on arXiv in July 2026: the method generates transparent solver components that an engineer can read, verify, and integrate into a standard workflow.

How the closed loop works

MILP-Evo iteratively generates candidate programs, loads them into the open-source solver SCIP via the PySCIPOpt library, and evaluates them based on real behavior on MILP problems. The result of each run becomes feedback that drives the further evolution of the code — with no human in the loop.

In the paper, the method was tested on the joint design of two solver components at once: the cut selector and the branching rule. These two components largely determine how quickly the solver finds an integer solution.

• Two components are jointly designed: the cut selector and the branching rule
• Candidates are executed directly in SCIP via PySCIPOpt — evaluation is based on the solver's end-to-end behavior, not a proxy metric
• Four evolution mechanisms: selection by outcome, targeted code repair, diagnostic reflection, and maintaining population diversity
• Testing was carried out on four benchmark families
• The output is readable white-box code, not an external "black-box" predictor

Why this beats ordinary ML for solvers

MILP-Evo produces explicit, readable code instead of an external predictor, and this is the main difference from previous data-driven approaches. As the authors put it, trained machine learning policies are typically represented as an opaque model that is hard to inspect, adapt, and deploy. Hand-written solver logic, on the other hand, is understandable and easy to integrate, but it is written by hand rather than derived from the solver's own feedback.

MILP-Evo tries to take the best of both worlds: the logic can still be read by eye, but it is synthesized not by a human but by an LLM-driven search loop.

"We study whether the automated design of MILP solver logic can be

framed as an LLM-guided closed-loop search over executable white-box components, evaluated directly on the solver's end-to-end behavior," the abstract of the MILP-Evo preprint on arXiv states.

How well does it work

Across four benchmark families, LLM-driven program evolution in a number of cases finds competitive policies specialized for a specific class of problems. The authors state their conclusion cautiously: the method delivers a competitive result "in several settings," not universal superiority over hand-crafted heuristics.

The mechanism itself matters more: the resulting components are ordinary code for SCIP that can be inspected, refined by hand, and deployed into the solver's standard workflow. This removes the main drawback of neural accelerators — the inability to understand and safely modify what the model has learned.

What this means

MILP-Evo is another example of how LLMs are turning from text generators into engines for automated code search: the model does not just predict — it writes executable components and improves them based on feedback from real runs. For optimization, this is a way to get heuristics that are both fast and transparent, without manual tuning for every class of problems.

Frequently Asked Questions

What is MILP, and why speed up its solvers?

MILP (mixed-integer linear programming) is a class of optimization problems with integer variables: logistics, scheduling, resource allocation. They are solved by solvers such as SCIP, and solving speed depends heavily on the heuristics for choosing cuts and branching — which is exactly what MILP-Evo designs.

How does MILP-Evo differ from neural solver accelerators?

Previous ML approaches represent a learned policy as an external opaque predictor that is hard to inspect and deploy. MILP-Evo instead produces explicit white-box code in PySCIPOpt that a person can read, modify, and integrate into SCIP's standard workflow.