Chain-of-Experts
Multi-agent operations-research pipeline that translates natural-language optimization problems into LP/MIP code and repairs from program execution feedback.
The loop
- Proposes
Specialized LLM roles extract mathematical terminology, formulate algebraic constraints, and emit Python PuLP/Gurobi solver code.
- Judged by
The Python interpreter and MIP solver execute the generated formulation and report syntax errors, infeasibility, or optimal objective values.
- Steers
A conductor chooses which role acts next; on a failed run the roles reflect on execution feedback in reverse order until one is at fault.
Headline result
Reported to significantly outperform state-of-the-art LLM-based approaches on the LPWP and ComplexOR operations research benchmarks.
What this does not prove
If the formulation specialist misinterprets an implicit constraint in the word problem without causing solver infeasibility, the MIP solver solves the wrong model.
How it generalizes the harness
An exact linear/integer programming solver computes the solution while program execution feedback drives reverse-order reflection by the roles.
Classification
| What it proposes | Formal specification |
|---|---|
| Generative model | GPT-4 multi-role formulation agents |
| Algorithmic authority | Gurobi / PuLP mixed-integer programming solver |
| Feedback | Verifier errors |
| What gets updated | Prompt context |
| When it runs | At inference time |
| Validity | Formal (machine-checked) |
| Evidence | Code released |
| Lineage | Original system |