Researchers from the SIA-NUS Digital Aviation Corporate Lab, together with Shanghai Jiao Tong University, Shanghai University of Finance and Economics, City University of Hong Kong and Cornell, developed LEAN-LLM-OPT, a framework that uses large language model agents to formulate large-scale optimisation models, and applied it to Singapore Airlines' (SIA) revenue management. The work was published on arXiv in January 2026 with code and data released on GitHub.
The problem is that building optimisation models for decisions such as fare capacity allocation is labour-intensive: an analyst must turn a business question into variables, constraints and coefficients drawn from several datasets. General-purpose LLMs fail at this when the data is large; in the paper, GPT-4.1 on its own could not produce a correct model for any SIA instance because it neither knew to pull values from the datasets nor had tools to do so.
LEAN-LLM-OPT splits the job across three agents. Given a query and its datasets, two upstream agents classify the problem and retrieve similar worked examples, then build a step-by-step workflow describing how such a model should be formulated. A downstream agent follows that workflow, calling custom data-handling tools to extract capacities, demand, prices and attraction values, and generates the final formulation and Python solver code. It was instantiated with GPT-4.1 and the open-source gpt-oss-20B.
For the SIA case study, the team used a booking simulator trained on SIA's real sales data to simulate a full 355-day selling season for a Singapore, Hong Kong and Manila network with Eco-lite and Eco-flexi fare types. They built a benchmark, Air-NRM, of 15 fare-capacity-allocation instances and 21 network-planning instances (joint seat allocation and flight selection) based on sales-based linear programming under a general attraction choice model.
LEAN-LLM-OPT with GPT-4.1 achieved 93.33% execution accuracy on capacity allocation and 92.06% on network planning, versus 80.00% and 85.71% for Gemini 3 Pro and 26.67% and 55.56% for GPT-5.2; the gpt-oss-20B version scored 86.67% and 82.54%. On network planning, its revenue optimality gap stayed below 14%. Ablations showed both parts matter: with tools but no workflow, accuracy fell to 20.00% and 30.16%, and with the workflow but no custom tools, to 0%.
The evaluation used simulated data derived from SIA sales; the paper does not report production use at SIA.