Watching Quantum Models Think: Hilbert-Space Interpretability in Quantum Transformer Blocks
Abstract
Deep learning models are powerful but opaque. As quantum machine learning matures, the field faces a defining choice: build quantum models that are equally opaque, or exploit the mathematical structure of quantum mechanics to make them inherently interpretable. We show that the latter is possible. By tracking quantum mutual information~(MI), entanglement entropy, and state fidelity through the layers of a Quantum Transformer Block (\qtb{}), a fully-coherent variational circuit with quantum analogues of both attention and feedforward, we gain direct insight into how the model processes information: which tokens it attends to, when correlations form, and why predictions fail. On four tasks with known dependency structure we show that (i)~learned MI matrices align with ground-truth task structure (AUC= 0.69 on lookup), (ii)~disabling entangling gates collapses accuracy from 100\% to 15\% while MI→ 0, proving entanglement is the mechanism, (iii)~accuracy and MI co-evolve during training (ρ = 0.92 on lookup), and (iv)~per-sample MI predicts prediction correctness on the conditional task with ROC AUC= 0.84. All results are validated on IBM Quantum hardware (ibm\_kingston, Heron~r2): the circuit's reasoning process, from product state through structured entanglement, is directly observable on a superconducting processor. These proof-of-concept results, obtained on small synthetic tasks, suggest that the physics of quantum computation can provide intrinsic interpretability signals with no direct classical counterpart, motivating study of whether this advantage persists at scale.
Related
- Inherent interpretability provides inherent value in quantum machine learning
- A Novel Steganography Scheme Using Quantum Hilbert Transform
- Metricity of separable quantum optimal transport with the Hilbert-Schmidt cost
- Variational study of the magnetization plateaus in the spin-1/2 kagome Heisenberg antiferromagnet: An approach from vision transformer neural quantum states
- All-fiber quantum current transformer based on nitrogen-vacancy centers
BibTeX
@misc{iacopetta2026watching,
title = {{Watching Quantum Models Think: Hilbert-Space Interpretability in Quantum Transformer Blocks}},
author = {Diego Iacopetta and Andrea Gasparini},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.23016},
eprint = {2609.23016},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.23016},
}
RIS
TY - UNPB
TI - Watching Quantum Models Think: Hilbert-Space Interpretability in Quantum Transformer Blocks
AU - Diego Iacopetta
AU - Andrea Gasparini
PY - 2026
DA - 2026/09/19
PB - arXiv
DO - 10.48550/arxiv.2609.23016
UR - https://arxiv.org/abs/2609.23016
AB - Deep learning models are powerful but opaque. As quantum machine learning matures, the field faces a defining choice: build quantum models that are equally opaque, or exploit the mathematical structure of quantum mechanics to make them inherently interpretable. We show that the latter is possible. By tracking quantum mutual information~(MI), entanglement entropy, and state fidelity through the layers of a Quantum Transformer Block (\qtb{}), a fully-coherent variational circuit with quantum analogues of both attention and feedforward, we gain direct insight into how the model processes information: which tokens it attends to, when correlations form, and why predictions fail. On four tasks with known dependency structure we show that (i)~learned MI matrices align with ground-truth task structure (AUC= 0.69 on lookup), (ii)~disabling entangling gates collapses accuracy from 100\% to 15\% while MI→ 0, proving entanglement is the mechanism, (iii)~accuracy and MI co-evolve during training (ρ = 0.92 on lookup), and (iv)~per-sample MI predicts prediction correctness on the conditional task with ROC AUC= 0.84. All results are validated on IBM Quantum hardware (ibm\_kingston, Heron~r2): the circuit's reasoning process, from product state through structured entanglement, is directly observable on a superconducting processor. These proof-of-concept results, obtained on small synthetic tasks, suggest that the physics of quantum computation can provide intrinsic interpretability signals with no direct classical counterpart, motivating study of whether this advantage persists at scale.
ER -