Overview
Scheduling industrial and logistics operations is a complex combinatorial optimization problem. A single instance can hold hundreds of thousands of variables that must be tuned to compute the optimal schedule, and even minor gains compound across multi-trillion dollar logistics networks into billions in savings. Yet logistics and scheduling remain in the grips of Good Old-Fashioned AI.
Industrial-162 collects 162 mixed-integer optimization problems spanning 22 domains, drawn from the MIPLIB 2017 reference library. Each system gets solver tools and a 600-second budget per problem; a problem counts as solved only when the formulation returns a valid optimal solution - or correctly proves the instance infeasible.
We formalized the logistics domains. Pramaana solves the full portfolio while frontier LLMs - even when equipped with solver tools - struggle to generate optimal valid solutions.
Key takeaways
- Frontier LLMs struggle to code optimization domains. Claude Fable 5, GPT-5.6 Sol, and Kimi K3 each solve fewer than 5% of the portfolio. Most fail to generate the right formulation code, and much of the generated code is bloated enough to blow the solver's 600-second timeout.
- Frontier capability is narrow and convergent. Kimi K3 and GPT-5.6 Sol pass the identical 4 problems out of 162; Claude Fable 5 passes a strict superset with 4 more. Capability looks limited to the small subset of problems encountered in training.
- Model sovereignty cuts into coverage. Claude Fable 5 - the leading LLM here - refused 7 biology-related problems as a biosafety risk. The filters appear too broad, missing the substance of the problems.
Benchmark scope
- Formulate the scheduling or allocation problem as a mixed-integer program
- Generate solver code and run it within a 600-second execution budget
- Return a valid optimal solution - or correctly prove the instance infeasible
- Cover 22 domains - logistics, manufacturing, telecom, energy, workforce, and more