OptiMind

Many real-world business problems can benefit from optimization, but translating decisions, constraints, and goals from natural language into mathematical models is slow. OptiMind is a small language model designed to convert business problems described in natural language into the mathematical formulations needed by optimization software. It is trained on a carefully curated, expert-aligned dataset and applies domain-specific hints and self-checks at inference time, improving its accuracy.

Figure 1: This figure illustrates how when given a new problem, OptiMind first classifies it into a category, such as scheduling or network design. It then applies expert hints that are relevant to that type of problem, which act as reminders to check for errors before generating a solution. For particularly challenging problems, the system generates multiple solutions and either chooses the one that appears most frequently or uses feedback to improve its response. This approach increases accuracy without requiring a larger model.

OptiMind incorporates knowledge from optimization experts both during training and when it’s being used to improve formulation accuracy at scale. It translates problems into the mathematical equations that optimization software can solve. This happens in three stages: improving training data quality with domain-specific hints, fine-tuning the model, and guiding its reasoning as it works.

OptiMind matches or exceeds the performance of much larger systems, can run locally to protect sensitive data, produces more reliable formulations, and reduces the time and expertise needed to prepare optimization models.

OptiMind is now available for experimental purposes on Microsoft Foundry.