A new presolve-ordering framework for mixed-integer programming solvers treats step sequencing as the thing to learn, not just the settings.
Researchers built ORDO (Operation-level Round-aware Dynamic Ordering), a system that treats MIP presolve as sequence generation rather than parameter tuning. Earlier learning-based presolve tools adjusted knobs but could not model that presolve steps are order-dependent - running the same steps in a different sequence can inflate the slowest solve times by several-fold. ORDO instead generates full action sequences autoregressively from a shared action vocabulary, then pairs that with what the team calls sequence racing: several candidate sequences run concurrently inside a modified build of the SCIP solver, and whichever one finishes is kept. To make that possible, the team added an execution-and-observation layer to SCIP that records exactly which actions ran and in which round.
Presolve - the cleanup phase before a solver actually starts crunching a mixed-integer program - routinely decides whether a problem is tractable or a time sink, and most prior automation treated it purely as a tuning problem. By making the order itself the learned object, ORDO reportedly generalizes to problem domains it never trained on, producing speedups with no retraining, which the authors say is a first for presolve action sequences. The fact that the size of the speedup varies by domain, and does not track how much training data existed for that domain, suggests it is picking up something structural rather than memorizing a corpus.
The biggest gains only show up once racing is switched on, so the honest read is promising, not solved - a fair description of most AI-for-optimization work right now.