Adaptive Policy-Guided Error Mitigation (APGEM): A Unified Learning-Based Framework for Noise Resilience in Quantum Machine Learning

Bisma Majid, Shabir Ahmad Sofi, Mir Mohammad YousufarXiv 2026 · arXiv

In the current "Noisy Intermediate-Scale Quantum" (NISQ) era, quantum computers are like experimental race cars: powerful, but prone to breaking down due to environmental "noise." To get useful results, researchers use Quantum Error Mitigation (QEM)—mathematical tricks to clean up noisy data. However, most researchers pick one mitigation technique at the start of an experiment and stick with it.

A new paper by Bisma Majid and colleagues argues that this "one-size-fits-all" approach is failing Quantum Machine Learning (QML). They propose APGEM, a framework that teaches a quantum system to choose its own error mitigation strategy on the fly.

TL;DR

APGEM is a "meta-learning" framework that uses a contextual-bandit controller to dynamically select the best error mitigation technique during a quantum algorithm's execution. By adapting to real-time feedback, it achieves 95.5% of the performance of a theoretical "perfect" selector, significantly outperforming static, hand-picked strategies.

What Problem Does This Paper Address?

Quantum hardware suffers from various types of noise: gates don't fire perfectly, qubits lose their state (decoherence), and measurements are often "fuzzy." While many QEM techniques exist—such as Zero-Noise Extrapolation (ZNE) or Probabilistic Error Cancellation (PEC)—they all have different costs and strengths.

The authors identify a critical gap: in QML, the quantum circuits and the noise they experience change as the model trains. A mitigation strategy that works well for a shallow circuit at the start of training might be too expensive or ineffective as the circuit deepens or the hardware drifts. Currently, there is no standard way to switch between these techniques automatically based on the actual state of the learning process.

Table 1: Summary of key literature on quantum error mitigation (QEM) relevant to Adaptive Policy-Guided Error Mitigation (APGEM).

Methods

The core of the paper is the Adaptive Policy-Guided Error Mitigation (APGEM) framework. It treats error mitigation as an online decision problem.

  1. The Controller: They use a "contextual bandit" (specifically the LinUCB algorithm). At each step, the controller looks at a "context vector"—a snapshot of the current state including noise rates, circuit depth, and training progress.
  2. The RBF Lift: Because the relationship between noise and the best mitigation technique isn't linear, they use a Radial-Basis Function (RBF) expansion to "lift" the context into a higher-dimensional space, making it easier for the linear controller to make smart decisions.
  3. The Mitigation Pool: The controller can choose from six "arms": No Mitigation, ZNE, PEC, Measurement Error Mitigation (MEM), Dynamical Decoupling (DD), and Circuit Optimization.
  4. Cost-Aware Utility: This is the "secret sauce." Instead of just chasing the highest accuracy, the reward function penalizes techniques that require too many "shots" (quantum measurements). It only picks an expensive technique like PEC if the accuracy gain is worth the massive increase in computational cost.
The APGEM decision loop: from noisy execution to feedback-driven technique selection.

Datasets

The researchers didn't use a single static dataset. Instead, they evaluated APGEM across:

  • Five QML Paradigms: Variational Classifiers (VQC), Kernel Methods (QSVM), Eigensolvers (VQE), Generative Models (QGen), and Reinforcement Learning (QRL).
  • 80 Noise Conditions: A factorial grid crossing different rates of depolarizing noise, readout error, and coherent phase error.
  • Real-World Validation: They used hardware-calibrated noise models derived from snapshots of eight real IBM quantum devices, ensuring the framework works on "messy" real-world noise, not just clean simulations.

Key Insights and Ideas

  • No Single Winner: The data confirms that no single fixed strategy wins across the board. For example, PEC is great for generative tasks, while "No Mitigation" is often the most rational choice for kernel methods where the cost of mitigation outweighs the accuracy gain.
  • Context over Algorithm: The authors found that the specific bandit algorithm (LinUCB vs. Thompson Sampling) mattered less than the fact that the controller was conditioned on context. Simply knowing the current noise and circuit state provides the biggest performance boost.
  • Noise-Awareness: The learned policy is physically intuitive. As depolarizing noise increases, the controller shifts from ZNE to the more powerful (but expensive) PEC. When noise becomes "severe" (beyond saving), it intelligently falls back to No Mitigation to save on costs.
  • The "Sampling Floor": A significant portion of the "error" in these systems isn't a failure of the policy, but irreducible "shot noise" from the quantum measurements themselves.

Key Results

  • Near-Oracle Performance: APGEM reached 95.5% of the utility of an "oracle" (a theoretical selector that always knows the best choice).
  • Hardware Success: On the real IBM device models, APGEM tracked the oracle almost exactly, achieving 99.9% of the possible utility.
  • Scalability: Crucially, the "wall clock" time and sampling cost for the controller to make a decision remained constant as they scaled from three to eight qubits. The overhead doesn't explode as the quantum computer gets bigger.
  • Statistical Significance: APGEM outperformed all six static strategies with high statistical confidence ($p \approx 1.1 \times 10^{-17}$).

Limitations and Future Work

While powerful, the authors note a few areas for growth:

  • Composition: Currently, APGEM picks one technique. In the future, it could learn to combine them (e.g., applying MEM and ZNE simultaneously).
  • Interaction Horizon: The controller's performance was still improving at 1,000 interactions. In very long-running experiments, the gains could be even higher.
  • Latency vs. Shots: The current utility function focuses on the "shot budget" (sampling cost). Future versions could be tuned for "hard latency" (real-time constraints) for different hardware architectures.

Summary updated

Abstract

Error mitigation stands between variational quantum algorithms and useful results on today's noisy hardware, yet in practice a mitigation method is usually chosen once, by hand, and applied unchanged for the life of an experiment. This is a poor fit for quantum machine learning (QML), where the circuit and the noise drift as training proceeds, so the best strategy in one regime is often wrong in another. We present a generalized form of Adaptive Policy-Guided Error Mitigation (APGEM), which treats the choice of mitigation technique as an online decision made from an algorithm's own feedback. A single contextual-bandit controller observes signals any variational algorithm already produces, such as fidelity, entropy, reward or loss, and gradient behavior, and selects among Zero-Noise Extrapolation, Probabilistic Error Cancellation, Measurement Error Mitigation, Dynamical Decoupling, and circuit optimization per step. Because it depends only on this shared feedback interface, the framework applies unchanged across supervised, unsupervised, reinforcement, and hybrid quantum-classical learning. Evaluated on five paradigms across 80 algorithm-and-noise conditions, adaptive selection reaches 95.5% of an oracle that always picks the best technique and significantly outperforms all six fixed strategies (Friedman $χ^{2}=92.18$, $p\approx1.1\times10^{-17}$; pairwise Wilcoxon $p<4\times10^{-6}$). The learned policy is noise-aware: its favored technique shifts as depolarizing and readout error rise. On noise models from eight real IBM devices, APGEM tracks the oracle almost exactly (99.9% of oracle utility) at a flat cost of roughly $1.3\times10^{4}$ shots from three to eight qubits.
BibTeX
@misc{majid2026adaptive,
  author = {Bisma Majid and Shabir Ahmad Sofi and Mir Mohammad Yousuf},
  title = {{Adaptive Policy-Guided Error Mitigation (APGEM): A Unified Learning-Based Framework for Noise Resilience in Quantum Machine Learning}},
  year = {2026},
  eprint = {2611.13815},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13815},
}
Figure