Enterprise AI Analysis
OptiRepair: Closed-Loop Diagnosis and Repair of Supply Chain Optimization Models with LLM Agents
This paper introduces OptiRepair, a novel two-phase framework for autonomously diagnosing and repairing infeasible supply chain optimization models using LLM agents. By combining domain-agnostic feasibility repair with domain-specific operational rationality validation, OptiRepair significantly outperforms frontier API models, highlighting the critical role of specialized training and explicit rationality checks in achieving reliable AI-driven operational planning.
Executive Impact
OptiRepair demonstrates a significant leap in AI's capability to automate complex operational tasks, dramatically reducing the need for scarce OR expertise in model maintenance and ensuring not just feasibility, but operational soundness in supply chain planning.
0%
Trained Model Rational Recovery Rate (RRR)
0%
Trained Model Phase 1 Feasibility Recovery
0%
Average API Phase 1 Recovery Rate
0pp
Performance Gap (Trained vs. Best API RRR)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Two-Phase Closed-Loop MDP
OptiRepair decomposes model repair into two phases: a domain-agnostic feasibility repair (Phase I) and a domain-specific operational rationality validation (Phase II). Phase I iteratively diagnoses and repairs constraint conflicts using solver IIS feedback. Phase II applies a domain oracle with five theory-grounded rationality checks to ensure the solution is operationally sound, not just mathematically feasible.
Error Classification & Rationality Checks
The system uses ten error types (ME-1 to ME-10) covering common multi-echelon inventory planning mistakes, grouped by operational subsystem. These range from easy-to-diagnose local conflicts (ME-3: Balance Violation) to hard-to-trace systemic issues (ME-1: Demand Inflation, ME-2: Lead Time Error). Operational rationality is enforced by five checks: Base-stock structure, Bullwhip ratio, Inventory allocation, Cost consistency, and Order smoothing, all grounded in classical supply chain theory.
Iterative STaR and GRPO Refinement
Two specialized 8B-parameter models are trained using iterative Self-Taught Reasoning (STAR) with solver-verified rewards and Group Relative Policy Optimization (GRPO). This approach allows the models to learn from solver feedback and refine their repair strategies. Phase I training focuses on feasibility restoration, while Phase II training ensures compliance with operational rationality criteria, connecting optimization to real-world operational decisions.
Benchmarking and Ablation Studies
Evaluated on 976 multi-echelon supply chain problems across 10 error types, OptiRepair's trained models achieved an 81.7% Rational Recovery Rate (RRR), significantly outperforming the best API models (42.2%). Ablation studies confirm that Phase 1 repair (solver interaction) is the critical bottleneck, with trained models achieving 97.2% recovery vs. 27.6% for APIs, while Phase 2 (operational rationality) pass rates are relatively high for APIs (76.3%).
81.7%
Rational Recovery Rate (RRR) achieved by trained 8B models, outperforming best API by 39.5 percentage points.
Enterprise Process Flow: OptiRepair Architecture
Infeasible LP
→
Phase I: Repair (LLM Agent + Solver IIS)
→
Optimal LP
→
Phase II: Validate (LLM Agent + Domain Oracle)
→
Rational Solution
| Metric | Trained 8B Pipeline | Best API Model (GPT-5.2) | Average API Model |
|---|---|---|---|
| Rational Recovery Rate (RRR) | 81.7% | 42.2% | 21.3% |
| Phase 1 Feasibility Recovery (RR) | 97.2% | 53.5% | 27.6% |
| Phase 2 Rationality Pass Rate (P2Pass) | 84.0% (Implied from RRR/RR) | 79.0% | 76.3% |
| Average Steps per Problem | 4.3 - 5.2 | 14.0 | 7.8 |
| Average Tokens per Problem | 3,180 - 3,439 | 20,730 | 30,078 |
The Solver Interaction Bottleneck
API models struggle significantly with Phase 1 repair, averaging only 27.6% Recovery Rate compared to 97.2% for trained models. This gap highlights that iterating on solver feedback and diagnosing IIS conflicts is a distinct and crucial skill that general LLMs lack without specialized training. In contrast, Phase 2 operational rationality checks (e.g., base-stock, bullwhip) are handled better by API models (76.3% pass rate), suggesting that domain knowledge, when explicitly codified as verifiable checks, is more accessible.
Calculate Your Potential AI ROI
Estimate the efficiency gains and cost savings your enterprise could achieve by integrating AI-powered optimization tools.
Estimated Annual Savings
Hours Reclaimed Annually
Your AI Implementation Roadmap
A typical journey to integrate OptiRepair-like capabilities into your enterprise operations.
Discovery & Pilot (Weeks 1-4)
Assess existing optimization models, identify common infeasibility patterns, and define key operational rationality criteria. Conduct a targeted pilot with a subset of models using OptiRepair's framework.
Customization & Training (Months 2-4)
Adapt the Phase II rationality oracle to your specific business rules and supply chain configurations. Fine-tune LLM agents on your proprietary model structures and historical repair logs, using solver-verified feedback.
Integration & Deployment (Months 5-6)
Integrate OptiRepair agents with your existing OR infrastructure (e.g., Gurobi, CPLEX). Deploy in a shadow mode for validation, then progressively roll out to production, starting with less critical models.
Monitoring & Iteration (Ongoing)
Continuously monitor agent performance, track repair efficiency and rationality compliance. Use feedback loops to retrain and refine models, incorporating new error types or operational requirements as your supply chain evolves.
Ready to Transform Your Operations?
Automate model maintenance, ensure operational rationality, and unlock the full potential of your supply chain optimization initiatives with enterprise AI. Our experts are ready to guide you.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat