Machine Learning Decoding of Circuit-Level Noise for Bivariate Bicycle Codes

The Big Picture

Imagine whispering a secret message across a crowded, noisy room. Every few feet, someone mishears a word and passes along a garbled version. You’d repeat the message in a clever pattern so that even if words get mangled, the original can be reconstructed.

Quantum computers face exactly this problem. The “noise” is quantum mechanical, the “words” are qubits (the quantum equivalent of binary bits, but far more fragile), and the stakes are whether quantum computers can ever reliably run complex calculations.

What doesn’t get enough attention is the classical decoder, a conventional software algorithm running in real time alongside the quantum processor. It watches measurement outcomes and infers which errors occurred. Without a fast, accurate decoder, error corrections pile up faster than they can be processed and the whole system grinds to a halt.

Researchers have spent years building good machine learning decoders for the most popular quantum error-correcting scheme, the surface code, which arranges qubits in a grid and detects errors using neighboring qubits. A newer, more promising class called quantum low-density parity-check (QLDPC) codes protects far more logical qubits with the same hardware, but has largely been out of reach for ML. That just changed.

A team at MIT has trained a recurrent, transformer-based neural network to decode circuit-level noise on Bivariate Bicycle (BB) codes, a family of QLDPC codes featured in a 2024 Nature paper by Bravyi et al. Their model doesn’t just match the best conventional decoder. It beats it, while running ten times faster in the worst case.

Key Insight: A machine learning decoder can outperform the leading classical algorithm for QLDPC codes in both accuracy and speed, putting real-time decoding for next-generation quantum computers within practical reach.

How It Works

After each round of syndrome extraction (measuring special operators on groups of qubits to detect errors without directly reading the qubits themselves), the decoder receives a string of 0s and 1s called a syndrome. Think of it as a fault report: it tells you that something went wrong, but not exactly what.

The decoder must infer which combination of physical errors most likely produced that syndrome, then output a correction. Get it wrong and you corrupt the logical qubit. Do it too slowly and the error backlog grows exponentially.

The best classical decoder for QLDPC codes is BP-OSD (Belief Propagation with Ordered Statistics Decoding). It propagates probability estimates across the code’s structure to identify the most likely errors. BP-OSD is powerful but has a fatal flaw: its runtime is wildly variable. Some syndromes decode in microseconds; others trigger worst-case cubic-time computations. For a quantum computer that needs corrections every clock cycle, unpredictable latency is a dealbreaker.

The MIT team’s approach combines three ideas:

• Code-aware self-attention: Standard transformers let every token attend to every other. Here, attention is restricted based on the BB code’s stabilizer generators, the measurement patterns that define how errors are detected. This constraint draws on classical coding theory, adapted for quantum codes, and it shrinks the search space considerably. Training converges faster and accuracy improves.

• Recurrent processing: Rather than ingesting all syndrome measurement rounds at once, the architecture processes one round at a time and carries a hidden state forward. This keeps the effective input size manageable as syndrome rounds accumulate.

• Autoregressive output: The model outputs a conditional probability distribution over logical errors, predicting each error type sequentially based on what it has already predicted. This approximates a maximum-likelihood decoder without requiring exponentially many output classes.

The decoder was trained and tested on two BB codes: the [[72,12,6]] code (72 physical qubits encoding 12 logical qubits with distance 6) and the larger [[144,12,12]] code. Training used circuit-level noise, the most realistic noise model available. Faulty measurements, gate errors, and error propagation between qubits during syndrome extraction are all part of the picture.

At a physical error rate of $p = 0.1%$ on the [[72,12,6]] code, the ML decoder achieved a logical error rate almost five times lower than BP-OSD. The speed story matters just as much. BP-OSD’s runtime has heavy tails, with occasional slow cases that could bottleneck a real quantum processor. The ML decoder runs in consistent, predictable time, an order of magnitude faster than BP-OSD’s worst cases.

Why It Matters

QLDPC codes are the most promising route to resource-efficient fault-tolerant quantum computing. The surface code devotes hundreds or thousands of physical qubits to protecting a single logical qubit. QLDPC codes like BB codes encode many more logical qubits per physical qubit; the [[72,12,6]] code protects 12 logical qubits with just 72 physical ones. That could slash hardware requirements by orders of magnitude.

IBM’s 2024 experimental demonstration of BB codes in Nature brought them from theory to hardware. The missing piece has been a decoder fast and accurate enough to use in practice.

Many assumed that ML decoding of QLDPC codes faced fundamental obstacles. The space of possible error patterns is enormous. The structure is non-local, unlike the surface code’s simple grid that convolutional networks exploit naturally. And the number of logical qubits is large. The right architecture, it turns out, can handle all three. Code-aware attention tackles non-locality, recurrence manages the growing syndrome history, and autoregressive outputs deal with the large logical qubit count.

The results are not yet definitive at all code sizes. On the [[144,12,12]] code, BP-OSD recovers its error-rate advantage at very low physical error rates. But the speed advantage holds across the board, and scaling to larger codes looks feasible.

Bottom Line: MIT researchers have built the first ML decoder to outperform BP-OSD on QLDPC codes in a practically relevant regime: 5× lower logical error rates and 10× faster worst-case runtimes on the [[72,12,6]] Bivariate Bicycle code. Real-time ML decoding for next-generation quantum processors now looks much more plausible.