3 Presentations

• 
  01:00 PM - 01:22 PM

  From Tweets to Decisions: Leveraging Large Language Models for Decision-Making

  • Alexandre Jacquillat, Massachusetts Institute of Technology
  • Shriya Karam, presenter, MIT

  Humanitarian logistics often rely on numerical data to calibrate decision-making models. However, these operations, such as disaster relief distribution, inherently involve multi-dimensional human experiences that numerical data alone may not capture. As a result, current operational efforts may fail to meet individual needs and respond effectively to evolving demand patterns. To address this gap, we propose a modeling framework that leverages qualitative data for data-driven decision-making in disaster relief, allowing agencies the ability to re-adjust operational plans in real-time. As a case study, we use Twitter data from Hurricane Harvey (2017) to enhance shelter location and relief distribution. Our approach employs large language models to parse hurricane severity tweets, integrating these insights into a multi-commodity flow and facility location problem to optimize resource allocation. Ultimately, our results show benefits in using qualitative data for proactive decision-making, as well as benefits in accounting for demand uncertainty based on predicted values.

• 
  01:22 PM - 01:44 PM

  OptiChat: Explaining Optimization Problems Using Large Language Models

  • Can Li, presenter, Purdue University
  • Hao Chen,
  • Gonzalo Constante Flores, Purdue University

  This talk introduces OptiChat, a chatbot powered by large language models, designed to explain optimization problems. A significant obstacle to the practical deployment of optimization models is the challenge associated with helping practitioners comprehend and interpret these models. OptiChat is capable of performing a range of tasks, including diagnosing infeasibilities, conducting sensitivity analyses, providing counterfactual explanations, and responding to general inquiries from users. OptiChat has also been tested on a dataset of different types of queries. We will end the talk by discussing the opportunities and limitations of the applications of LLM in optimization.

• 
  01:44 PM - 02:06 PM

  Bridging the Gap between Natural Language and Constraint Models: An Overview of Ner4Opt and Text2Zinc

  • Serdar Kadioglu, presenter, Brown University

  Optimization Technology has made significant theoretical and practical advances, yet a key challenge remains: the cognitive barrier of translating problem descriptions into formal constraint models. In parallel, the rise of Large Language Models (LLMs) has revolutionized human-computer interaction, enabling communication across multiple modalities. However, despite the impressive progress in LLMs, accurately generating constraint models from free-form natural language descriptions continues to be a difficult task.

  To address this gap, we present two complementary contributions: i) Ner4Opt, a specialized named entity recognition system tailored for optimization problems, and ii) Text2Zinc, a dataset and evaluation framework for assessing LLM-powered approaches to generating MiniZinc models from natural language descriptions. We also discuss the open challenges and future directions in making decision-making tools more accessible to non-expert users.