IBM Research claims a proven separation between shallow quantum circuits and large language models, in a paper posted to arXiv on August 4.
Two problem classes are involved. The first asks a system to compute a correct output for a given input, which is roughly what answering a prompt requires. The second asks it to draw samples from a target probability distribution. On one problem of each kind, the paper shows shallow quantum circuits holding an advantage no classical model of this type can reach, and the proof carries no assumption that the classical side is weak.
“Separating quantum circuits from classical LLMs” runs 60 pages and carries six figures. Srinivasan Arunachalam, Arkopal Dutt, Hari Krovi and Rik Sengupta are its authors. IBM’s write-up of the result appeared on September 15, with Ryan Mandelbaum contributing.
Depth is the variable that makes this work. A shallow circuit holds that depth fixed even as the qubit count grows, which is what sets the class apart from machines built to brute-force search at scale.
The lineage matters here. IBM researchers placed a related separation in Science back in 2018, and that gap has been widened against stronger classical models ever since. What the new paper changes is the target, which is language models rather than abstract circuits.
The authors are explicit about the limits. Their findings are theoretical. Today’s quantum machines still fight error rates, while the systems they outrun sit on hardware built over years of industrial investment.
