An AI learned to compile quantum circuits by reading them as images
MIT and IBM turned a quantum chip's underlying math into a single picture and taught a 3 billion parameter model to translate it back into the gates a quantum chip runs.
MIT and IBM turned a quantum chip's underlying math into a single picture and taught a 3 billion parameter model to translate it back into the gates a quantum chip runs.
MIT and IBM have built a language model that reads a quantum circuit's underlying math like a photograph and writes back the sequence of gates that runs on a real quantum chip. The result, from an MIT and IBM team, was reported in a paper at the IEEE Quantum Week / QCE 2026 conference, and treats a continuous-valued quantum matrix as a native image input rather than converting it into symbolic code first.
The team's paper, "Aligning Quantum Operators with Large Language Models" on arXiv as 2606.13811v1, puts a number on how well the approach works. Its writers report 99.4% compilation success on 4-qubit circuit synthesis under best-of-80 sampling, and 91% compliance with text-form constraints on a first attempt. Best-of-80 means the model samples 80 candidate gate sequences and keeps the best one; it is not a first-shot success rate. Quantum Computing Report's coverage of the paper carries the same metrics.
A quantum circuit is, mathematically, a unitary operator: a square matrix that describes exactly what the chip does. The MIT-IBM system rewrites that unitary as a real-valued 256x256 Pauli Transfer Matrix, the standard object physicists use to compare how close two quantum operations are, and feeds that matrix to the language model as a single-channel grayscale image. The image is split into 16x16 patches, projected through a small linear layer and a two-layer MLP, and dropped into the same token embedding space the model uses for words.
At each step the system reads what is left of the matrix, predicts one small π/8-Pauli rotation gate in reverse execution order, and peels that gate off by multiplying its inverse back onto the residual. It stops when the residual is small enough that the rebuilt circuit matches the original. Output is one gate at a time, with the input matrix shrinking as decoding goes.
The backbone is IBM's Granite 4.0 Micro, a 3-billion-parameter hybrid enterprise model fine-tuned with standard supervised next-token training, the same recipe used to adapt language models to new tasks in other domains. Earlier AI-for-quantum work has almost always forced the model to read a symbolic proxy: OpenQASM source, gate names, Qiskit-style Python. Reading a continuous-valued matrix as an image is the architectural move the paper claims as new.
4-qubit Clifford+T synthesis, the system's actual job, is a small but expressive gate set used as a benchmark, run inside a 256-way Pauli-rotation basis. A 3-billion-parameter model is small by 2026 standards. The metrics are from the authors' own evaluation and have not yet been independently reproduced. This is a conference paper, not a product, and the camera-ready version on the IEEE Quantum Week record should be reconciled with the arXiv preprint before the headline numbers are treated as locked.
Continuous scientific math that today has to be hand-coded into symbolic text can, in principle, be image-ified so any large language model with a visual pathway can read and edit it. Whether the same recipe generalizes past 4 qubits, past first-attempt success, and out of the conference-paper tier is the open question. The paper is on the program at IEEE Quantum Week 2026.