This is the official open-source repository accompanying the paper “little m: An AI Agent for Industrial Process Optimization.” It provides the IPC-Bench evaluation benchmark, a reference implementation of the little m workflow, and the domain knowledge base used for retrieval.
IPC-Bench is the multimodal benchmark introduced in the paper for evaluating industrial process optimization modeling. It contains 50 canonical, textbook-derived scenarios from process control and chemical engineering optimization.
Each case combines a natural-language process description with any available process diagram and asks the system to formulate a structured mathematical optimization model. The benchmark therefore evaluates both semantic interpretation and grounding in process structure, rather than text-only equation generation.
data/
├── 1.md – 50.md # 50 problem descriptions with ground truth models
└── figures/ # Process flow diagrams and schematics (41 images)
└── {id}_{source-case}_{figure-index}.png
Each markdown file ({id}.md) contains a structured optimization problem with:
- Problem Description — Narrative background describing the industrial process, operational goals, and physical context, with an embedded figure reference.
- Variables — Decision variables with mathematical symbols and physical descriptions.
- Objective Function — The optimization target (
min/max) in LaTeX. - Constraints — Physical, operational, and safety constraints in LaTeX, including system dynamics, bounds, and conservation laws.
The reference release consists of two files:
| File | Description |
|---|---|
Agent_EN.yml |
The Dify implementation of the paper's non-interactive workflow: information structuring, task planning, retrieval enhancement, knowledge retrieval, strategy determination, and mathematical modeling. |
industrial_optimization_knowledgebase.md |
The knowledge base used by the retrieval stage. It contains 45 general industrial pathways and 10 brewing-specific pathways organized around symptoms, optimization objectives, candidate algorithms, required information, mathematical patterns, and applicability limits. |
To reproduce the reference implementation in Dify, upload industrial_optimization_knowledgebase.md as a knowledge base, import Agent_EN.yml, configure the required model providers, and bind the uploaded knowledge base to the retrieval node with the same name. Knowledge-base IDs are specific to each Dify instance, so the binding must be configured after import.
@inproceedings{ye2026little,
title = {{little} m: An AI Agent for Industrial Process Optimization},
author = {Ye, Yongchao and He, Xinyu and Boshoff, Dutliff and Kuo, Way and Li, Lishuai},
booktitle = {Findings of the 2026 Conference on Empirical Methods in Natural Language Processing, EMNLP 2026},
year = {2026},
address = {Budapest, Hungary},
}