City University of Hong Kong researchers have introduced little m, an AI assistant that turns process descriptions and diagrams into mathematical models for engineering review. In their September 15 preprint, experts preferred its formulations over two selected language-model baselines, offering evidence for assistance with an early step in process optimization.
That step is deciding exactly what an optimization problem means. The released benchmark separates each answer into quantities the engineer can adjust, an objective to minimize or maximize, and equations or limits the process must respect. These are called decision variables, the objective function and constraints. A process diagram adds equipment connections that a written description alone may leave unclear.
The public workflow organizes the supplied information before choosing a strategy and drafting equations. A retrieval stage searches a curated engineering knowledge base for relevant methods and information requirements. The implementation assigns the main reasoning stages to Gemini 2.5 Pro. Engineers can inspect its prompts, retrieved guidance and intermediate outputs alongside the final answer.
The knowledge base shows how that guidance works through a brewing example. One entry addresses cooling heat recovered at the wrong time for later hot-water demand. It calls for information about temperatures, tank inventory and batch schedules, then proposes energy balances and scheduling methods subject to equipment and quality limits. This is reusable modeling guidance in the library; the example is not a reported brewery deployment or measured energy saving.
The evaluation used textbook-derived problems. Eight specialists reviewed 20 cases with system identities hidden; little m received 66% of overall-quality preference votes against Qwen3-Next-80B-A3B-Instruct and DeepSeek-V3.2. Automated comparison across 50 cases favored its variable and constraint formulations, while DeepSeek led on the objective-function metric. These are comparisons within the authors' test, not an industry-wide ranking or proof that every component caused an improvement.
The release gives engineers material to examine that narrower result. Its IPC-Bench dataset, short for Industrial Process Control Benchmark, includes problem statements and reference models. A workflow file for the Dify application and a separate knowledge-base document expose the reference implementation. Reproducing it requires configuring model providers and connecting the knowledge base; the repository describes this released workflow as non-interactive, so readers should distinguish it from the paper's simulated clarification experiments.
The study leaves numerical feasibility and real plant performance untested, and prior training exposure to textbook cases cannot be excluded. Its present value is a candidate model that an engineer can inspect before implementation. The next decisive evidence would be successful solver execution and validation on independently collected industrial cases.