⚛️ quantum physics

Turning Quantum Compiler Configuration Selection Upside Down: A Top-Down Paradigm

This paper challenges the conventional bottom-up approach to quantum compiler configuration by proposing a top-down paradigm that leverages an exhaustively searched benchmark (QuOPTBench) to derive transferable rules and a rule-guided ML framework (QuTOP), enabling fast and high-quality configuration selection without iterative online search.

Original authors: Ming Zhong, Xiangyu Ren, Kangyu Zheng, Zhi Zeng, Shenggen Zheng, Shaohua Li, Zhiding Liang

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ming Zhong, Xiangyu Ren, Kangyu Zheng, Zhi Zeng, Shenggen Zheng, Shaohua Li, Zhiding Liang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Quantum computers promise to solve problems that are currently impossible for today's machines, from designing new medicines to cracking complex codes. However, these machines are incredibly fragile. Their internal components, called qubits, are prone to errors caused by even the slightest heat or vibration. To make a quantum computer work, scientists must translate a user's problem into a sequence of instructions the machine can understand, a process known as compilation. But because the hardware is so noisy, the way these instructions are arranged matters immensely. A slight change in the order of operations or the specific path the data takes can mean the difference between a result that is accurate and one that is pure noise. The challenge for researchers is finding the perfect set of instructions for every single problem on every specific machine, a task that requires balancing the quirks of the hardware against the structure of the software.

For years, the standard approach to solving this puzzle has been to start with a guess and then try to improve it step by step. Imagine a traveler trying to find the best route through a mountain range by starting at the bottom and climbing higher, checking the view at each step to see if they are getting closer to the peak. In the world of quantum computing, this "bottom-up" method involves picking a computer, trying a specific set of settings, seeing how well it works, and then tweaking the settings again and again. This process is slow and often gets stuck on a local peak, missing the true summit because the initial choice of mountain was wrong. A team of researchers at The Chinese University of Hong Kong decided to turn this entire process upside down. Instead of starting at the bottom and climbing, they looked at the entire mountain range from above, mapped out every possible route in advance, and then used those maps to guide future travelers instantly.

The researchers began by building a massive reference library called QuOptBench. They took 268 different quantum problems and, for each one, they tested every single combination of available computers and software settings that could possibly work. There were 960 different combinations for each problem. By running all of these tests, they created a perfect "oracle"—a record of the absolute best setting for every single scenario. This exhaustive search took thousands of hours of computer time, but it provided a complete map of the terrain. With this map in hand, they could see exactly which settings worked best for which types of problems and which machines, revealing patterns that the old step-by-step methods had missed.

When the team first tried to use advanced artificial intelligence models to predict the best settings directly, the results were disappointing. These models, which are trained to be helpful and thorough, tended to choose the most complex and aggressive settings available, assuming that "more is better." They would pick the most powerful computer and the most intense optimization settings, even when a simpler, quieter approach would have produced a better result. The researchers realized that these models were overthinking the problem, favoring complexity over the specific needs of the task. To fix this, they used a powerful reasoning model to study the perfect maps they had created. They asked the model to look at 68 specific examples where the standard approach failed and the perfect approach succeeded, and to explain why. From this analysis, the model extracted 13 clear, transferable rules. These rules were not vague suggestions but specific instructions, such as "if the problem is small and the computer is large, use a simple layout" or "if the problem involves many connections, choose a computer with better wiring."

Using these distilled rules, the team built a new system called QCompass. Instead of guessing and checking, QCompass looks at a new problem, checks it against the 13 rules, and immediately narrows down the search to the most promising options. It then uses lightweight mathematical models to quickly score these options and pick the winner. This approach is dramatically faster than the old methods. While the traditional step-by-step search could take nearly 48 minutes to find a good solution for a single problem in specific baseline configurations (such as the Device-Aware BO method), QCompass completes its selection process in just about 3.3 seconds on average. More importantly, it finds better solutions. In their tests, QCompass consistently outperformed the old methods, especially for difficult problems where there was a lot of room for improvement. It also proved to be more reliable when faced with new types of computers and problems it had never seen before, showing that the rules it learned were truly general.

The study suggests that the future of tuning quantum software may not lie in complex, endless searches, but in learning from past successes to make smart, instant decisions. By shifting from a bottom-up search to a top-down selection based on clear, extracted knowledge, the researchers have shown that it is possible to navigate the complex landscape of quantum computing with greater speed and accuracy. Their work indicates that while the hardware remains difficult, the software that controls it can be made much more effective by understanding the underlying patterns of what works and what does not. This top-down paradigm offers a promising path forward, turning a slow, trial-and-error process into a fast, knowledge-driven one.

Technical Summary: Turning Quantum Compiler Configuration Selection Upside Down

1. Problem Definition

Quantum compiler configuration selection involves two tightly coupled decisions: device selection (choosing a target quantum backend) and option selection (configuring compiler passes such as layout, routing, and optimization levels). The goal is to maximize the execution fidelity of a compiled quantum circuit.

Existing state-of-the-art approaches, such as MQT Predictor, typically employ a bottom-up search paradigm. These methods first select a device and then iteratively tune compiler options using reinforcement learning (RL) or heuristic search. The paper argues that this strategy is fundamentally limited because a suboptimal initial device choice can impose a "fidelity ceiling" that subsequent option tuning cannot overcome. Furthermore, while classical compiler auto-tuning often faces prohibitively large search spaces, practical quantum compiler configuration spaces (e.g., in Qiskit) are bounded and discrete, making exhaustive offline evaluation feasible.

2. Methodology

The authors propose a top-down paradigm that shifts from iterative online search to deriving reusable configuration knowledge from offline oracle results. The methodology consists of three main components:

A. QuOptBench: An Oracle Benchmark

The authors constructed QuOptBench, the first exhaustive benchmark for Qiskit compiler configuration selection.

Scope: It covers 268 quantum circuits from QASMBench and VeriQBench.
Search Space: For each circuit, the study exhaustively enumerated 960 device-option configurations across four fake backends (16, 27, 65, and 127 qubits) and various optimization levels ( $O0$ – $O3$ ), layout, routing, and translation methods.
Oracle: The "oracle" configuration for each circuit is defined as the one yielding the highest effective estimated fidelity ( $F_{eff}$ ), a normalized metric that accounts for gate errors and circuit size.
Findings: Exhaustive search revealed that 213 cases (79.5%) have at least one capacity-feasible device whose best achievable fidelity (across all option combinations) is lower than the default $O3$ configuration fidelity on another feasible device. This confirms that an unfavorable device choice can impose a fidelity ceiling that subsequent option selection cannot overcome.

B. Analysis of LLM-Based Direct Selection

The paper evaluated ten Large Language Models (LLMs) as a natural realization of direct configuration selection.

Result: Vanilla LLMs consistently underperformed the oracle and often performed worse than a simple baseline ( $O3$ on the best device).
Cause: LLMs exhibited a strong bias toward configuration aggressiveness, preferring complex settings (e.g., $O3$ , Sabre routing, synthesis) regardless of circuit or device characteristics. They failed to recognize scenarios where simpler, less aggressive configurations yield higher fidelity.

C. Rule Distillation and QCompass

To overcome LLM limitations and the cost of exhaustive search, the authors developed a rule-guided framework:

Rule Distillation: Using Claude-Opus-5, the authors analyzed 68 sampled circuits where the oracle differed significantly from a baseline. They extracted 13 transferable rules governing device and option selection based on circuit features (e.g., qubit count, entanglement ratio, parallelism) and device characteristics (e.g., connectivity, error rates).
QCompass Framework: A lightweight, rule-guided Machine Learning framework designed for fast inference.
- Predictive Models: It trains four Extremely Randomized Trees (Extra Trees) models to predict optimization opportunity, fidelity gain, timeout risk, and compilation time.
- Policy Gate: Based on predicted optimization headroom, QCompass selects a policy: KEEP (retain baseline), SAFE (local modifications), or SEEK (broader exploration).
- Rule-Guided Candidate Generation: The distilled rules constrain the candidate space, filtering out unlikely configurations before scoring.
- Scoring: A unified objective function scores candidates based on predicted fidelity gain, penalizing timeout risk and compilation time.

3. Key Contributions

QuOptBench: The first exhaustive benchmark providing empirical oracle configurations for 268 circuits, enabling evaluation against the theoretical best rather than just standard baselines.
Empirical Study of LLMs: Demonstrated that general-purpose LLMs alone cannot reliably reproduce oracle-quality configurations due to a bias toward overly aggressive settings.
Rule Distillation Workflow: A novel process to extract transferable, human-interpretable rules from oracle decisions, contrasting $O3$ configurations with exhaustive-search oracles.
QCompass: A compact, rule-guided ML framework that achieves fast configuration selection without iterative online search, shifting the paradigm from bottom-up tuning to top-down selection.

4. Experimental Results

The evaluation compared QCompass against bottom-up methods (RL, Bayesian Optimization), nearest-neighbor retrieval, and direct LLM selection across three difficulty levels (defined by optimization headroom).

Performance vs. Baselines:
- Level 1 (Low Headroom): QCompass matched the baseline ( $O3$ on best device) in 67/76 cases with no statistically significant difference, avoiding the degradation seen in bottom-up methods that make unnecessary changes.
- Level 2 & 3 (High Headroom): QCompass significantly outperformed all baselines. On Level 3, it achieved a 28.56% improvement in effective estimated fidelity over the baseline, compared to ~20% for the best LLM and ~24% for Joint RL.
Efficiency: QCompass reduced selection latency to ~3.3 seconds per circuit. This is orders of magnitude faster than Bayesian Optimization (~~2,800s) and significantly faster than LLMs (~~17–110s), while maintaining superior solution quality.
Generalization: When tested on new circuits and devices (not seen during training), QCompass maintained a 9.06% average improvement over the baseline, whereas bottom-up RL methods yielded negative improvements, highlighting better generalization of the top-down approach.
Noise-Aware Evaluation: Evaluations using noisy simulation (Hellinger fidelity) confirmed that QCompass's advantages hold under realistic noise conditions, achieving a 5.71% mean improvement.

5. Significance and Claims

The paper claims that transferring knowledge from offline oracle characterization offers a promising alternative to conventional bottom-up quantum compiler tuning.

Paradigm Shift: It challenges the assumption that iterative search is necessary for configuration selection, demonstrating that the bounded nature of quantum compiler spaces allows for the extraction of high-quality, transferable rules.
Practicality: By combining distilled rules with lightweight predictive models, QCompass achieves the best overall balance between solution quality and selection latency, making it suitable for practical deployment where iterative search is too slow and LLM reasoning is too expensive or unreliable.
Limitations: The authors modestly acknowledge that their results are based on specific benchmarks (QASMBench, VeriQBench), fake backends, and Qiskit V2.4.1. They note that generalization to substantially different workloads, hardware architectures, or real hardware with calibration drift remains future work.

In conclusion, the paper establishes that a top-down, rule-guided approach can effectively navigate the quantum compiler configuration space, outperforming both iterative search and direct LLM inference in terms of fidelity and efficiency.