Back to feed
arXiv cs.AI·

Beyond Objective Equivalence: Constraint Injection for LLM-Based Optimization Modeling on Vehicle Routing Problems

Signal
78
Hype
25
In three linesVRPCoder, an 8B model, translates natural-language vehicle routing scenarios into Gurobi code. Authors propose constraint injection to verify constraints are neither silently omitted nor spuriously added. With GRPO, VRPCoder reaches 93% Pass@1, outperforming Claude-Sonnet-4.5 by 28 points on VRP benchmarks.
Read source
Your take?
Code generationReinforcement learningBenchmarksEvals

Summary generated by Claude — human-verified