Go to ICLR 2026 Conference homepageMIND: Market Interpretation DSL for Unified Market Design and Simulation
Keywords: Market Design, Domain-Specific Language, DSL, AI Copilot, Auctions, Matching Markets, Mechanism Design, Code Generation, Symbolic AI, Computational Economics
Abstract: Market mechanisms such as auctions and matchings coordinate supply and demand at scale, yet their implementations remain locked in rigid procedural code that hinders iteration and auditing. We introduce the Market Interpretation DSL (MIND), a typed language and toolchain for declarative market specification to achieve unified market design and simulation. MIND comprises (i) a core grammar with a phased Intermediate Representation (IR) and economic safety checks, (ii) a natural language assistant that translates descriptions into DSL with automated diagnostics and safe rewrites, and (iii) rule-based simulation and convex optimization backends. Using synthetic specifications generated across 87 domains with held-out validation, our fine-tuned Llama-3-8B assistant achieves 96.33% semantic correctness, measured as IR equivalence to gold programs, surpassing few-shot GPT-4o at 91.41%. Across second-price auctions, multi-stage auctions, and matching markets, MIND reduces specification complexity by approximately 79% in lines of code compared to Python implementations. In a preregistered within-subjects study with 17 participants, mechanism modifications were completed 4 to 10 times faster using MIND. Code, dataset, and models will be released upon acceptance.
Primary Area: applications to computer vision, audio, language, and other modalities
Submission Number: 19233
Loading