World's first Transformer-based quantum error mitigation validated on real IBM Quantum hardware
| Version | Model | Backend | Noisy | Mitigated | Improvement |
|---|---|---|---|---|---|
| v1.0 | Standard QNN | ibm_marrakesh 3q | 98.5% | 99.1% | +0.6% |
| v2.0 | Upgraded QNN | ibm_fez 5q | 88.99% | 93.06% | +4.57% |
| v3.0 | Transformer QNN | ibm_fez 5q | 56.90% | 99.80% | +75.40% |
| Circuit | Noisy | Standard QNN | Transformer QNN |
|---|---|---|---|
| 1 | 85.66% | 99.30% | 99.46% |
| 2 | 84.92% | 98.78% | 99.64% |
| 3 | 83.14% | 99.62% | 99.41% |
| 4 | 81.96% | 98.15% | 99.43% |
| 5 | 82.65% | 98.00% | 99.41% |
| AVG | 83.67% | 98.77% | 99.47% |
🏆 Transformer QNN wins on ALL 5/5 circuits 🔬 Average improvement: +18.89% over raw hardware
World's first Transformer-based quantum error mitigation system validated on real IBM Quantum hardware. Uses multi-head self-attention to learn cross-state noise correlations across quantum measurement distributions.
Standard QNN treats each measurement independently. Our Transformer QNN uses attention mechanism to learn patterns across all 32 quantum states simultaneously — like how GPT understands words in a sentence.
- IBM Quantum (ibm_fez 156-qubit, ibm_marrakesh 156-qubit)
- Qiskit + Qiskit Runtime
- PyTorch Transformer (204,225 parameters)
- Python 3.12
- 5 qubits, 4 layers, full entanglement
- Circuit depth: 235 gates
- 4096 measurement shots
- Input: 32 noisy state probabilities
- 4 Transformer blocks
- 4 attention heads per block
- Embedding dimension: 64
- Output: 32 corrected probabilities
pip install -r requirements.txt
python quantum_project.py| File | Description |
|---|---|
quantum_project.py |
Main v2.0 code |
transformer_qnn_results.png |
Transformer benchmark plot |
qnn_results_v2.png |
v2.0 results plot |
benchmark_results.json |
5-circuit benchmark data |
requirements.txt |
Dependencies |
If you use this work please cite:
Cheela, A. (2026). Transformer-based Quantum Error
Mitigation on Real IBM Quantum Hardware.
GitHub: github.com/cheelaakhil/quantum-error-mitigation
Akhil Cheela github.com/cheelaakhil

