Solver-Guided GRPO Lifts LLM Formulation Selection From 21.45% to 50.42% Hit Ratio@1 on JD.com Warehouse Allocation
A paper posted 2026-07-28 (arXiv 2607.25956) treats multi-warehouse inventory allocation as instance-wise operations-research formulation selection: rather than writing one mixed-integer program, an LLM picks from a library of solver-executable MIP formulations, each encoding a different allocation priority, based on the instance's demand concentration, inventory imbalance, and forecast volatility. Training goes from expert-conditioned SFT for schema learning, to margin-weighted IPO preferences derived from solver-evaluated allocation-quality gaps, to GRPO with per-instance expert-score reward lookup. On instances from JD.com, GRPO raises Hit Ratio@1 from 21.45% to 50.42% and Hit Ratio@2 from 70.47% to 82.31%, delivering a 12.57 percentage point allocation-accuracy gain over the incumbent and narrowing the gap to the ex-post oracle to 4.85 points.
↳ Follow the thread