LLM-Guided Initialization for Accelerated Hybrid Quantum-Classical Medical Image Classification

Technical Summary: LLM-Guided Initialization for Accelerated Hybrid Quantum-Classical Medical Image Classification

Problem Statement
Variational Quantum Algorithms (VQAs) face a critical barrier to practical utility known as the "barren plateau" phenomenon. In this scenario, the variance of the cost function gradient decays exponentially as the number of qubits or circuit depth increases. This exponential suppression renders gradient-based optimization infeasible for deep circuits, as the optimizer receives vanishingly small signals to navigate the loss landscape. While strategies such as local cost functions and layerwise training exist, determining optimal initial parameters for hardware-efficient ansatze remains an open challenge. Random initialization often places the optimizer in regions of parameter space where gradients are effectively zero.

Methodology
This paper evaluates AdaInit (Adaptive Initialization), a framework proposed by Zhuang and Cunningham that utilizes Large Language Models (LLMs) to generate initial parameter configurations for Parameterized Quantum Circuits (PQCs). The study investigates a simplified, single-query variant of AdaInit, contrasting it with the iterative refinement used in the original proposal.

The experimental setup involves:

• Task: Binary classification (benign vs. malignant) on the DMR-IR mammography dataset (infrared thermography images).
• Architecture: A hybrid quantum-classical model featuring a classical convolutional feature extractor reducing images to 8 dimensions, followed by a 4-qubit, 2-layer Variational Quantum Circuit (VQC) with 16 total parameters. The circuit uses $R_X$ and $R_Y$ rotations and CNOT entangling gates.
• Initialization Strategies:
  1. Random Initialization: Parameters sampled uniformly from $[-0.5, 0.5]$.
  2. AdaInit (LLM-Guided): Parameters generated by querying the Qwen2.5-3B-Instruct model with a structured prompt describing the circuit architecture and the objective of maximizing gradient variance. The LLM proposes values typically in the range $[-3, 3]$ with structured inter-parameter correlations.
  3. Identity Block Initialization: A deterministic baseline where angles are fixed to $\pi/2$.
• Simulation: Experiments were conducted using NVIDIA CUDA-Q for GPU-accelerated quantum simulation on an NVIDIA L40S GPU.

Key Contributions

1. Empirical Validation in Medical Imaging: The paper validates the AdaInit framework on a real-world medical image classification task, demonstrating its applicability beyond theoretical benchmarks.
2. Performance Metrics: The study reports a 14.6× higher gradient variance at initialization for AdaInit compared to random initialization (0.0095 vs. 0.0006). This translates to a 160× training speedup, reducing convergence time from 176 seconds to 1.1 seconds, while maintaining identical final classification accuracy (61.4%).
3. Barren Plateau Verification: The authors systematically verified the barren plateau phenomenon by measuring gradient variance decay across circuit depths (2 to 8 layers). They observed a 64% decay in gradient variance from depth 2 to depth 8, confirming the exponential suppression predicted by theory and motivating the need for informed initialization.
4. Parameter Analysis: The paper analyzes the characteristics of LLM-proposed parameters, noting that they exhibit magnitude diversity (spanning $[-3, 3]$), sign alternation, and inter-layer correlations. These features prevent the circuit from approximating a Haar-random unitary, thereby preserving gradient information.

Results and Analysis

• Gradient Variance: The primary finding is that LLM-guided initialization places the optimizer in "trainable regions" of the parameter space where gradient signals are strong. The 14.6× improvement in initial variance directly correlates with the 160× reduction in training time.
• Convergence vs. Accuracy: Both initialization methods converged to the same final accuracy of 61.4%. The authors attribute this to the shallow nature of the 2-layer circuit, where barren plateaus are mild and the loss landscape likely contains a single dominant basin of attraction. The speedup is attributed to the optimizer making meaningful progress from the first iteration due to larger effective parameter updates, rather than an improvement in the model's final expressibility.
• Classical Baseline: A classical Convolutional Neural Network (CNN) achieved 100% accuracy, highlighting the current limitations of NISQ-era quantum models (4 qubits, 16 parameters) in terms of expressibility compared to classical architectures. The study explicitly states its goal is not to claim quantum advantage in accuracy, but to demonstrate training acceleration.

Significance and Claims
The paper claims that a single LLM query is sufficient to yield informative parameters without the need for iterative refinement, offering a low-overhead path to improved trainability. The findings validate AdaInit as a compatible strategy for GPU-accelerated quantum backends (specifically CUDA-Q).

The authors position this work as a demonstration of "foundation model-assisted quantum computing," where knowledge embedded in LLMs (trained on scientific literature) is transferred to guide quantum algorithm design. They suggest that while the current results are modest in terms of circuit depth, the approach is essential for scaling to deeper circuits where random initialization would result in vanishing gradients. The study concludes that AdaInit effectively mitigates the barren plateau problem at initialization, enabling practical speedups in hybrid quantum-classical training workflows.