ORLM
Operations Research Language Model
Trains open-weight LLMs on synthetic operations-research formulations whose programs execute in the COPT mixed-integer solver, then solves problems via LLM+Solver.
The loop
- Proposes
An LLM translates complex industrial scheduling, routing, and supply-chain descriptions into formal mathematical models and COPT Python code.
- Judged by
The COPT mathematical programming solver executes the generated MIP/LP script and reports the solver optimum.
- Steers
During synthetic data curation (OR-Instruct) unexecutable programs are discarded; at inference the code is run and the solver optimum is reported.
Headline result
Outperformed GPT-4 across NL4Opt, MAMO, and IndustryOR benchmarks using a 7B open-weight model paired with the COPT solver.
What this does not prove
As in all Formalize-and-Solve systems, solver optimality applies to the generated algebraic model rather than unstated domain nuances.
How it generalizes the harness
An exact mathematical programming solver (COPT) both discards unexecutable synthetic training programs and computes the final optimization solution.
Classification
| What it proposes | Formal specification |
|---|---|
| Generative model | Fine-tuned Mistral-7B / LLaMA-3 ORLM |
| Algorithmic authority | COPT mixed-integer and linear programming solver |
| Feedback | Verifier errors |
| What gets updated | Model weights and training data |
| When it runs | At inference time |
| Validity | Formal (machine-checked) |
| Evidence | Code released |
| Lineage | Original system |