little m is an AI agent designed for industrial process optimization. It bridges a domain‑specific knowledge base with large language model (LLM) interaction to translate messy natural‑language specifications and process diagrams into rigorous mathematical models. The main workflow consists of:
- Parsing textual requirements and spatial diagrams to extract variables and physical relations;
- Retrieving corresponding continuous multi‑physics equations from the dedicated knowledge repository;
- Using the LLM to generate constraints, objective functions, and variable declarations while guaranteeing correct syntax;
- Emitting a well‑formed optimization model.
A typical output model looks like
$$ \min{x}\; \sum{i} c_i x_i \ \text{s.t.}\; A x = b,\; x \ge 0 $$
little m is systematically evaluated on the newly introduced Industrial Process Control Benchmark (IPC‑Bench), which comprises 50 canonical scenarios that require joint reasoning over text and process diagrams. Automated structural checks and double‑blind human evaluation show that little m substantially outperforms state‑of‑the‑art LLMs in producing semantically correct models. The evaluation measures formulation quality rather than solver feasibility or closed‑loop industrial performance.
Review