RL-SPH: feasible routing and scheduling plans, no ILP solver
KAIST's ICML method hits 100% feasibility across five benchmarks and scales 67x beyond its training data — a weekend swap for a pricey commercial solver.
Copy markdown
Feasible plans, no solver in the loop
RL-SPH reports a 100% feasibility rate across five benchmarks and cuts the average primal gap 28.6x, producing usable integer-program solutions without handing off to a commercial solver like Gurobi or SCIP.
It scales far past its training set
The method finds a first feasible solution about 2.5x faster and held up on problems up to 67x larger than its training data. Training the model itself takes roughly 30 minutes, 14.7x faster than prior learned heuristics.
Build this weekend: a solver-free scheduler
If your work touches vehicle routing, staff rostering, or factory production scheduling, this points at prototyping an optimizer with no commercial solver license attached. The paper, arXiv 2411.19517, is the starting point.
The catch: a paper, not a package
There is no confirmed drop-in library yet. RL-SPH is an ICML 2026 result you would implement yourself, and you should validate it against your own constraints before trusting it in production.
Elsewhere: Copilot is dropping two Gemini models
GitHub is retiring Gemini 2.5 Pro and Gemini 3 Flash from Copilot (announced July 31). If your Copilot setup pins either model, switch before they are pulled.