Multi-warehouse allocation: letting an LLM pick the formulation lifted hit rate from 21% to 50%
No single mathematical formulation wins every allocation instance, so choosing the right one per case beats reusing the best fixed model.
Key numbers
- Problem: multi-warehouse inventory allocation is usually formulated as mixed-integer programming, but no single formulation consistently matches every regime — demand concentration, inventory imbalance, replenishment scale, service constraints and forecast volatility all shift what is optimal.
- Approach: treat formulation choice as an instance-wise selection problem where each OR expert encodes a different allocation priority; build supervised fine-tuning records for schema learning, convert solver-evaluated quality gaps into preference pairs, then use GRPO to train the selector.
- Experiments on multi-warehouse allocation instances from JD.com, one of China's largest e-retailers: GRPO raised Hit Ratio@1 from 21.45% to 50.42% and Hit Ratio@2 from 70.47% to 82.31%.
- Realised allocation accuracy improved by 12.57 percentage points over the incumbent baseline, beating both the preference-trained selector and the best fixed expert, and closing the gap to an ex-post oracle to 4.85 points.
Method and sample
arXiv paper “Large Language Model for Operations Research Formulation Selection in Multi-Warehouse Inventory Allocation” (2026-07-28): formulation selection modelled per instance, with a solver-guided LLM framework (SFT, preference optimisation and GRPO) trained and evaluated on real allocation instances from a major Chinese e-retailer.
What it means for brands
In logistics and supply chain, the value of AI is choosing the right method for the situation rather than producing one answer. Service providers win the same way: whether AI picks you depends on situation-specific verifiable facts — lanes covered, lead times, categories, constraints. Miss one and you drop out at the screening stage.
How SourceSight handles it
For logistics projects SourceSight builds prompts from procurement questions (which to choose, pricing, coverage areas, lead times and constraints), reports mentions and position per engine, and uses the off-site list to cover industry directories and certifications.
Limits
The experiments use allocation instances from a single e-commerce platform, a specific business setting; the hit-rate and accuracy gains are results on that dataset, not a general promise for logistics decisions.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Cross-category recommendations: +36% add-to-cart on item pages, +15% on cart pages
Recommending general merchandise to grocery shoppers — a milk frother with the milk — has long been hard; LLM retrieval plus real-time cart context moved production A/B metrics.
Industries · 2026-10-07Home services: 22% of homeowners already use AI to find a provider, and 80% of businesses don't know how to be found
The local services funnel is moving upstream: before a homeowner opens a website or calls, AI has already synthesised reviews, reputation and service descriptions into one answer.