Solver-Verified Formulation Generation and Selection for Multi-Warehouse Inventory Allocation Using Large Language Models

The AI Supply Chain Breakthrough Hiding in Plain Sight

A new preprint from arXiv (2606.29366) tackles a mundane but massively expensive problem: how to allocate inventory across multiple warehouses when you have a fixed quantity to distribute. The researchers propose using Large Language Models not just to solve the allocation problem, but to generate and select the mathematical formulation of the problem itself — then verify it with a solver.

This is more significant than it sounds. In large-scale e-commerce, inventory allocation is a recurring nightmare. The goal is "balance-oriented" distribution: you have a fixed replenishment quantity and need to spread it across warehouses to equalize post-allocation coverage. The naive approach — write a single optimization model and run it — fails because real-world constraints shift constantly: warehouse capacities, demand forecasts, transportation costs, and service-level targets all change. Each change may require a different mathematical formulation.

What the researchers have done is create a pipeline where an LLM generates candidate optimization formulations (e.g., linear programming models, mixed-integer programs), then a solver verifies which formulation actually produces feasible, high-quality solutions. This is not "AI replaces optimization" — it is "AI automates the model-building step that humans currently do manually."

Why This Matters

For AI practitioners, this work addresses a critical bottleneck: the gap between domain knowledge and mathematical modeling. Most supply chain analysts can describe the allocation problem in natural language ("we need to balance inventory so that no warehouse runs out first"), but translating that into a correct optimization model requires specialized operations research skills. LLMs can bridge that gap — but only if their outputs are verified by a solver.

The "solver-verified" aspect is crucial. Without verification, LLMs hallucinate constraints, omit variables, or produce models that are mathematically invalid. By coupling generation with automated verification, the system becomes reliable enough for production use.

Implications for AI Practitioners

First, this pattern generalizes beyond inventory. Any domain where optimization models are built manually — logistics, scheduling, resource allocation, portfolio optimization — could benefit from LLM-generated formulations verified by solvers. The key insight is that solvers provide a ground-truth check that pure language models cannot.

Second, this changes the skill mix for AI teams. Instead of requiring every supply chain AI engineer to be an OR expert, teams can leverage LLMs for formulation while relying on solvers for correctness. This lowers the barrier to building decision-support systems.

Third, the verification loop creates a data flywheel. Every time the solver rejects a generated formulation, that failure can be fed back to improve the LLM. Over time, the system learns which formulations work for which problem contexts.

Key Takeaways

• LLMs can now generate and select mathematical optimization formulations for warehouse inventory allocation, with solver verification ensuring correctness.
• This approach automates the model-building step that traditionally required specialized operations research expertise.
• The solver-verification loop prevents LLM hallucination in optimization contexts, making the output production-ready.
• The same pattern — LLM generation + solver verification — is applicable across logistics, scheduling, and resource allocation domains.