New AI Turns Your Plain-English Problems Into Solvable Math

Some of the hardest problems in business have no obvious answer: which delivery truck takes which route, how to pack a warehouse, how to schedule nurses so everyone gets fair hours. Mathematicians turn these puzzles into a special recipe called QUBO (short for "quadratic unconstrained binary optimization") — essentially a math format that special solvers, including quantum computers, can chew on.

The problem is the translation step. Turning a messy real-world situation into a correct QUBO recipe takes specialist knowledge, careful naming of every yes/no decision, and penalty weights that stop the computer from cheating. Researchers Niloy Kumar Mondal and Md Rizwan Parvez built an AI system that does this translation automatically from a plain-language description. Instead of one AI answering in a single shot, several AI helpers work together and then check and fix their own mistakes in loops. That self-repair step, they found, mattered more than anything else.

To test it, the team built QUBOBench: 100 optimization problems drawn from 12 industries and published research. Their system solved 68% correctly, a 22% jump over asking a single AI once. That is meaningful progress but far from perfect — roughly one in three problems still came out wrong, and a wrong formulation gives a confidently wrong answer. A human expert would still need to check the math before anyone routes real trucks with it.

The code and data are free and open-sourced, so other researchers and companies can build on it. Nobody is shipping this as a product yet, and the jump from academic test problems to messy real-world logistics is a big one. Still, it points to a future where the scarce skill — knowing how to phrase a business problem for a solver — becomes something you can just type out and describe.