IQM published results on August 1, 2026, from a technique it calls Quantum-Informed Surrogate Sampling, or QISS, run on its 54-qubit Emerald processor. The headline comparison: a QISS setup using a 3-layer (p=3) QAOA circuit reportedly outperformed a standard 17-layer (p=17) QAOA circuit on MaxCut, on average. If you've read our QAOA tutorial, that gap should stand out immediately. Circuit depth is one of the main things that kills near-term quantum algorithms on noisy hardware, since every added layer adds more opportunities for error. A method that gets QAOA-level results at a fraction of the depth is worth understanding properly rather than skimming past.
What QISS does
The core idea is a shift in what the quantum circuit is for. Standard QAOA tries to directly sample the answer: run a deep circuit, measure, and hope the output distribution is concentrated on good solutions. QISS instead uses a shallow circuit to generate statistical data, specifically low-order correlators, that a classical optimizer then uses to search for a solution. IQM's framing is that a shallow circuit's effective capacity exceeds a much deeper one once you add this kind of classical post-processing, because the quantum hardware's job shrinks to producing informative statistics rather than the entire answer. The method requires only O(N) low-order correlators, meaning the amount of quantum-generated data needed scales linearly with the problem size rather than growing combinatorially.
Why shallower matters more than the MaxCut number itself
MaxCut is the standard toy benchmark for this class of algorithm, useful for controlled comparison but not itself the point of the result. The point is what shallow circuits open up on real, noisy hardware. Deeper circuits accumulate gate errors and decoherence faster than today's NISQ-era error correction compensates for, which is the same problem our error mitigation guide covers from a different angle, and the same reason shot-reduction techniques matter for variational algorithms generally. A technique that reaches deep-circuit-quality output from a shallow circuit is attacking the noise problem structurally instead of patching around it after the fact.
What to check before taking the comparison at face value
"Outperforms on average" invites the obvious question: average over what problem sizes, and by what margin? IQM's own release frames this as noise-resilience validation on real hardware rather than a simulated result, which is a meaningfully stronger claim than a purely theoretical proposal, but the comparison is still against vanilla QAOA specifically, not against every classical or hybrid solver. Our quantum machine learning reality check makes the general point that near-term quantum advantage claims need a genuine classical baseline, not only a same-hardware quantum comparison, to mean much. Whether QISS beats the best classical MaxCut heuristics on the same instances is the harder and more useful question, and it isn't the one this result answers.
What to watch next
Two things would move this from "promising technique validated on one benchmark" to "practically useful": results on problem instances beyond MaxCut, particularly the kind of real-world graph and portfolio problems covered in our quantum portfolio optimization business case, and a head-to-head comparison against strong classical solvers rather than only against deeper QAOA circuits.