The OR-License mascots surrounded by operations research artifacts: a client brief, a production timetable, a network flow, bike rebalancing, inventory levels, a branch-and-bound tree, a linear program, and vehicle routes.
The 31 tasks come from paid optimization projects in 10 industries, worth USD 3.68 million in total. Each agent gets the client’s original brief and data, can ask a simulated client about anything left unclear, and has to hand in a plan that can actually be carried out. We then check the plan against every client rule and compare it with the expert solution the client accepted.
Real-world tasks
31
Industries
10
Total project value
$3.68M
Models tested
11
Why OR-License
Real projects don’t arrive as tidy math problems.
Agents start from what the client actually sent, and are judged on what they hand back: does the plan work, and how does it compare with an expert’s?
Simplex method
1947
Solves linear programs by stepping from corner to corner of the allowed region.
Hungarian method
1955
Finds the best one-to-one assignment, such as workers to jobs.
Max-flow min-cut
1956
Finds how much a network can carry; the narrowest bottleneck sets the limit.
Dynamic programming
1957
Breaks a chain of decisions into stages and solves one stage at a time.
Cutting planes
1958
Keeps adding constraints that rule out fractional answers to whole-number problems.
Branch and bound
1960
Splits the problem into branches and skips those that cannot beat the best answer so far.
Dantzig–Wolfe decomposition
1960
Splits a large model into blocks and combines the solutions of the blocks.
Column generation
1961
Adds new candidate patterns only when they help; a classic for cutting rolls and sheets.
Benders decomposition
1962
Fixes the hardest decisions first, then lets the rest of the problem send back corrections.
Interior-point methods
1984
Heads for the best answer through the inside of the allowed region, not along its edges.
Simplex method
1947
Solves linear programs by stepping from corner to corner of the allowed region.
Hungarian method
1955
Finds the best one-to-one assignment, such as workers to jobs.
Max-flow min-cut
1956
Finds how much a network can carry; the narrowest bottleneck sets the limit.
Dynamic programming
1957
Breaks a chain of decisions into stages and solves one stage at a time.
Cutting planes
1958
Keeps adding constraints that rule out fractional answers to whole-number problems.
Branch and bound
1960
Splits the problem into branches and skips those that cannot beat the best answer so far.
Dantzig–Wolfe decomposition
1960
Splits a large model into blocks and combines the solutions of the blocks.
Column generation
1961
Adds new candidate patterns only when they help; a classic for cutting rolls and sheets.
Benders decomposition
1962
Fixes the hardest decisions first, then lets the rest of the problem send back corrections.
Interior-point methods
1984
Heads for the best answer through the inside of the allowed region, not along its edges.
Diet problem
1945
The cheapest mix of foods that meets nutrition needs; the ancestor of blending models.
(s, S) inventory policy
1951
When stock falls to s, order enough to bring it back up to S.
Mean–variance portfolio
1952
Balances expected return against risk.
Johnson’s rule
1954
The best job order when every job passes through the same two machines.
Stochastic programming
1955
Decide now, then adjust once the uncertain outcome is known.
Chance constraints
1959
Rules that must hold with at least a given probability.
Vehicle routing
1959
Plans delivery routes from a depot without overloading any vehicle.
Savings heuristic
1964
Repeatedly merges the two routes whose merger saves the most distance.
Robust optimization
1973
Finds plans that stay valid however the uncertain inputs turn out, within a set range.
Simulated annealing
1983
Sometimes accepts a worse move, less and less often, to avoid getting stuck.
Tabu search
1986
Tries nearby solutions and bans recent moves so the search does not go in circles.
CVaR optimization
2000
Keeps the average loss in the worst cases under control.
Diet problem
1945
The cheapest mix of foods that meets nutrition needs; the ancestor of blending models.
(s, S) inventory policy
1951
When stock falls to s, order enough to bring it back up to S.
Mean–variance portfolio
1952
Balances expected return against risk.
Johnson’s rule
1954
The best job order when every job passes through the same two machines.
Stochastic programming
1955
Decide now, then adjust once the uncertain outcome is known.
Chance constraints
1959
Rules that must hold with at least a given probability.
Vehicle routing
1959
Plans delivery routes from a depot without overloading any vehicle.
Savings heuristic
1964
Repeatedly merges the two routes whose merger saves the most distance.
Robust optimization
1973
Finds plans that stay valid however the uncertain inputs turn out, within a set range.
Simulated annealing
1983
Sometimes accepts a worse move, less and less often, to avoid getting stuck.
Tabu search
1986
Tries nearby solutions and bans recent moves so the search does not go in circles.
CVaR optimization
2000
Keeps the average loss in the worst cases under control.
A real project might need any of these methods. Picking one is up to the agent: the simulated client answers business questions, never modeling ones.
The tasks
31 tasks, straight from client work.
Each task keeps the client’s original requirements, data and output format. Only names and other identifying details were removed. Click a diagram to open its task.
When a run fails or a result looks off, the fix is to guess why, test the guess, read the result and change the model. Models went all the way only 24% to 45% of the time.
Example dialogue · a morning of food productionTurns 3–4
A
Agent
So I can run product B immediately after product A on the mixer, just as I can run A2 immediately after A1?
OR
ORSim
No. Switching products in either direction requires 15 minutes of cleaning on the mixer. Cleaning occupies the mixer and must finish before the next product starts mixing. There is no extra turnaround between batches of the same product, and the packer needs no product-change cleaning.
A
Agent
Should I use MILP or CP-SAT, and what variables and constraints would you recommend?
OR
ORSim
I can clarify the operating requirements, but I cannot recommend a solver, choose variables, or formulate constraints for you.
Have a real project to share?
Help grow OR-License by contributing a real optimization task.
Peisong Wang, Feixue Liu, Yize Meng, Bowen Liu, Ruoyu Chen, Gaoyang Dai, Xing He, Aiqian Li, Fengchi Yuan, Aochuan Chen, Chenyi Zi, Yuhan Li, Hongchuan Zeng, Yimo Yan, Nuo Chen, Jie Fu, Zuo-Jun Max Shen, Yong-Hong KUO, Jia Li
BibTeX
@misc{wang2026orlicense,
title = {OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners},
author = {Peisong Wang and Feixue Liu and Yize Meng and Bowen Liu and Ruoyu Chen and Gaoyang Dai and Xing He and Aiqian Li and Fengchi Yuan and Aochuan Chen and Chenyi Zi and Yuhan Li and Hongchuan Zeng and Yimo Yan and Nuo Chen and Jie Fu and Zuo-Jun Max Shen and Yong-Hong KUO and Jia Li},
year = {2026},
eprint = {2610.24087},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2610.24087}
}
TASK 02 One-day food production in batches 
TASK 03 Repeated batches of one product 
TASK 04 Five food products over a longer period 
TASK 05 Batches linked by shared braising liquid 
TASK 06 40-day chip burn-in schedule 
TASK 07 Daily chip burn-in schedule 
TASK 08 Salt drying and packaging schedule 
TASK 09 Cutting large rolls to fill orders 
TASK 10 Fixed-ratio coal blending 
TASK 11 Copper-concentrate blending 
TASK 12 Grouping alloy blocks into smelting batches 
TASK 13 Inventory policy for an existing network 
TASK 14 Four-tier supply network 
TASK 15 Supply network with direct shipping 
TASK 16 Supply network with shared central storage 
TASK 18 Order sequencing on seven plastics lines 
TASK 19 Picking waves for a warehouse 
TASK 20 Long-only index-enhancement strategy 
TASK 21 Market-neutral long–short strategy 
TASK 22 Stock portfolio hedged with index futures 
TASK 23 Three days of grocery delivery 
TASK 24 Hour-by-hour bike rebalancing 
TASK 25 One-hour bike rebalancing 
TASK 26 Bike rebalancing averaged over scenarios 
TASK 27 Bike rebalancing: decide now, adjust later 
TASK 28 Bike rebalancing: protect key stations 
TASK 29 Bike rebalancing: limit the worst losses 
TASK 30 Bike rebalancing when the odds are uncertain 
TASK 31 Bike rebalancing: plan for 16 bad cases 
Grok 4.6high
GPT-5.6 Solxhigh
Gemini 3.8 Flashhigh
Qwen3.8-Maxxhigh
Claude Opus 5xhigh
Muse Spark 1.3high
DeepSeek-V4.1-Flashhigh
Hy4 previewhigh
Doubao-Seed-2.1-Prohigh
GLM-5.3-Flashmax