Teaching a quantum decoder to work with missing photon signals

Quantum error correction creates a classical computing problem alongside the quantum one: interpreting a stream of imperfect measurements quickly enough to guide subsequent operations. An accurate answer that arrives too late can slow the machine.

In a new preprint, researchers affiliated with PsiQuantum and AMD introduce PRISM, a transformer-based decoder designed for photonic quantum computing. Their simulations show accuracy comparable to strong conventional decoders across several error-correction protocols. A hardware-oriented version also offers a route to short, predictable processing times. However, the paper’s fastest headline timing is a projection supported by implementations of individual layers, not a completed full-network implementation at that speed.

Missing information can still be useful

In fusion-based quantum computing, small entangled states are connected through joint measurements called fusions. Their outcomes supply information needed both to perform the computation and to interpret its errors.

Photons can be lost, and fusion measurements can fail probabilistically. These events produce erasures: outcomes that are unavailable or unreliable at known locations. Knowing where information is missing helps the decoder, even though it does not reveal the missing value.

Other faults can flip outcomes without announcing their locations. Redundant checks reveal inconsistencies, collectively called a syndrome. The decoder combines that syndrome with the erasure pattern to infer how errors have affected the encoded, or logical, information. Corrections can then be tracked classically rather than immediately applied as physical operations.

The challenge is to use those clues accurately without letting difficult error patterns cause unpredictable processing delays.

A transformer constrained by the code

PRISM receives two kinds of input: check results and flags identifying erased outcomes. It represents checks and possible error mechanisms as nodes in a graph, connecting them wherever a fault would affect a check.

Its attention mechanism exchanges information only along these connections. Unlike unrestricted attention, this sparse structure avoids comparing every input with every other input. At fixed model width and depth, the authors report that memory and computation grow linearly with protocol size. That scaling statement does not guarantee that a fixed-size model will retain its accuracy as problems grow.

The network then predicts whether logical information has been flipped. It examines several mathematically equivalent representations of each logical quantity, aligns their predictions using the observed syndrome, and combines their scores.

Training starts with easier, lower-noise examples before increasing the noise. The authors found that starting with strong noise could leave the model predicting essentially uninformative 50–50 answers.

Accuracy under a specified noise model

The benchmarks cover fusion networks associated with surface codes, color codes and bivariate bicycle codes, a family of quantum low-density parity-check codes. Some protocols reach distance 24. Here, distance describes the minimum number of faults that can produce an undetected logical error—not simply the number of qubits.

The simulations use a simplified model: each fusion outcome is independently erased with probability x, while a non-erased outcome flips with probability x/10. This captures an erasure-dominated setting, but not all correlations or imperfections of an operating photonic device.

The authors report that PRISM matches the reference decoders across the tested noise range and performs better at lower error rates. The references differ by protocol: matching-based decoding for one network, integer-programming-based decoding for two others, and belief propagation with post-processing for the bicycle-code networks. These are not all exact optimal decoders.

For one distance-24 linked-loopy-diamond network, PRISM produced no logical failures in 30 billion simulated trials at an erasure probability of 1.2%. That is a finite-sample result, not evidence that its failure probability is zero.

What the microsecond figures establish

For hardware deployment, the team developed Q-PRISM, using ternary matrix weights and predominantly two-bit core activations, alongside simplified arithmetic. Some intermediate quantities retain higher precision. The compact models sacrifice accuracy relative to the full-precision reference decoders.

The hardware study targets a distance-8 linked-loopy-diamond block spanning eight code cycles. The complete accelerator was mapped onto an AMD VP1902 field-programmable gate array at 250 MHz. Its reported block latency is about 10.1 microseconds, with a steady-state interval of 2.39 microseconds between blocks.

Individual message-passing and readout layers met timing at 400 MHz. From these results, the authors project a full-block latency of 6.59 microseconds and a 1.56-microsecond interval between blocks. Completing the full mapping at 400 MHz remains future work. The interval is shorter than the latency because different blocks overlap within the processing pipeline. Reported latency includes input loading.

The study therefore supports a specific route toward fast decoding, not a demonstration of a complete fault-tolerant photonic computer. Its benchmarks use periodic blocks; connecting such blocks into a larger decoding system requires buffering and communication work left unresolved here. Extending the hardware result to larger distances, realistic device noise and a running quantum processor remains an important test.