Interpretable and scalable quantum natural language processing
September 18, 2024
The central question that pre-occupies our team has been:
“How can quantum structures and quantum computers contribute to the effectiveness of AI?”
In previous work we have made notable advances in answering this question, and this article is based on our most recent work in the new papers [, ], and most notably the experiment in [].
This article is one of a series that we will be publishing alongside further advances – advances that are accelerated by access to the most powerful quantum computers available.
Large language Models (LLMs) such as ChatGPT are having an impact on society across many walks of life. However, as users have become more familiar with this new technology, they have also become increasingly aware of deep-seated and systemic problems that come with AI systems built around LLM’s.
The primary problem with LLMs is that nobody knows how they work - as inscrutable “black boxes” they aren’t “interpretable”, meaning we can’t reliably or efficiently control or predict their behavior. This is unacceptable in many situations. In addition, Modern LLMs are incredibly expensive to build and run, costing serious – and potentially unsustainable –amounts of power to train and use. This is why more and more organizations, governments, and regulators are insisting on solutions.
But how can we find these solutions, when we don’t fully understand what we are dealing with now?1
At ԹϺ, we have been working on natural language processing (NLP) using quantum computers for some time now. We are excited to have recently carried out experiments [] which demonstrate not only how it is possible to train a model for a quantum computer in a scalable manner, but also how to do this in a way that is interpretable for us. Moreover, we have promising theoretical indications of the usefulness of quantum computers for interpretable NLP [].
In order to better understand why this could be the case, one needs to understand the ways in which meanings compose together throughout a story or narrative. Our work towards capturing them in a new model of language, which we call DisCoCirc, is reported on extensively in this .
In new work referred to in this article, we embrace “compositional interpretability” as proposed in [] as a solution to the problems that plague current AI. In brief, compositional interpretability boils down to being able to assign a human friendly meaning, such as natural language, to the components of a model, and then being able to understand how they fit together2.
A problem currently inherent to quantum machine learning is that of being able to train at scale. We avoid this by making use of “compositional generalization”. This means we train small, on classical computers, and then at test time evaluate much larger examples on a quantum computer. There now exist quantum computers which are impossible to simulate classically. To train models for such computers, it seems that compositional generalization currently provides the only credible path.
1. Text as circuits
DisCoCirc is a circuit-based model for natural language that turns arbitrary text into “text circuits” [, , ]. When we say that arbitrary text becomes ‘text-circuits’ we are converting the lines of text, which live in one dimension, into text-circuits which live in two-dimensions. These dimensions are the entities of the text versus the events in time.
To see how that works, consider the following story. In the beginning there is Alex and Beau. Alex meets Beau. Later, Chris shows up, and Beau marries Chris. Alex then kicks Beau.
The content of this story can be represented as the following circuit:
Figure 1. A text circuit for a simple story, involving three actors Alex, Beau andChris, who have a number of interactions with one another, making up a story –the circuit is to be read from top to bottom.
2. From text circuits to quantum circuits
Such a text circuit represents how the ‘actors’ in it interact with each other, and how their states evolve by doing so. Initially, we know nothing about Alex and Beau. Once Alex meets Beau, we know something about Alex and Beau’s interaction, then Beau marries Chris, and then Alex kicks Beau, so we know quite a bit more about all three, and in particular, how they relate to each other.
Let’s now take those circuits to be quantum circuits.
In the last section we will elaborate more why this could be a very good choice. For now it’s ok to understand that we simply follow the current paradigm of using vectors for meanings, in exactly the same way that this works in LLMs. Moreover, if we then also want to faithfully represent the compositional structure in language3, we can rely on theorem 5.49 from our book Picturing Quantum Processes, which informally can be stated as follows:
If the manner in which meanings of words (represented by vectors) compose obeys linguistic structure, then those vectors compose in exactly the same way as quantum systems compose.4
In short, a quantum implementation enables us to embrace compositional interpretability, as defined in our recent paper [].
3. Text circuits on our quantum computer
So, what have we done? And what does it mean?
We implemented a “question-answering” experiment on our ԹϺ quantum computers, for text circuits as described above. We know from our new paper [] that this is very hard to do on a classical computer due to the fact that as the size of the texts get bigger they very quickly become unrealistic to even try to do this on a classical computer, however powerful it might be. This is worth emphasizing. The experiment we have completed would scale exponentially using classical computers – to the point where the approach becomes intractable.
The experiment consisted of teaching (or training) the quantum computer to answer a question about a story, where both the story and question are presented as text-circuits. To test our model, we created longer stories in the same style as those used in training and questioned these. In our experiment, our stories were about people moving around, and we questioned the quantum computer about who was moving in the same direction at the end of the stories. A harder alternative one could imagine, would be having a murder mystery story and then asking the computer who was the murderer.
And remember - the training in our experiment constitutes the assigning of quantum states and gates to words that occur in the text.
Figure 2. The question-answering task for the language of text circuits as implementable on a quantum computer from []. Above the dotted line is the text we consider. Below are upside-down text circuits which constitute the question we ask. The boxes with words are parameterized as quantum gates. The diagram on the left constitutes one possible answer to the question, and the one on the right the other. Can you figure out what the text is and what the questions are?
4. Compositional generalization
The major reason for our excitement is that the training of our circuits enjoys compositional generalization. That is, we can do the training on small-scale ordinary computers, and do the testing, or asking the important questions, on quantum computers that can operate in ways not possible classically. Figure 4 shows how, despite only being trained on stories with up to 8 actors, the test accuracy remains high, even for much longer stories involving up to 30 actors.
Training large circuits directly in quantum machine learning, leads to difficulties which in many cases undo the potential advantage. Critically - compositional generalization allows us to bypass these issues.
Figure 3. A simplified plot from [] showing that increasing the sizes of circuits when testing doesn’t affect the accuracy, after training small-scale on ordinary computers. The number of actors correlates with the text size. H1-1 is the name of the ԹϺ quantum computer that was used.
5. Real-world comparison: ChatGPT
We can compare the results of our experiment on a quantum computer, to the success of a classical LLM ChatGPT (GPT-4) when asked the same questions.
What we are considering here is a story about a collection of characters that walk in a number of different directions, and sometimes follow each other. These are just some initial test examples, but it does show that this kind of reasoning is not particularly easy for LLMs.
The input to ChatGPT was:
What we got from ChatGPT:
Can you see where ChatGPT went wrong?
ChatGPT’s score (in terms of accuracy) oscillated around 50% (equivalent to random guessing). Our text circuits consistently outperformed ChatGPT on these tasks. Future work in this area would involve looking at prompt engineering – for example how the phrasing of the instructions can affect the output, and therefore the overall score.
Of course, we note that ChatGPT and other LLM’s will issue new versions that may or may not be marginally better with ‘question-answering’ tasks, and we also note that our own work may become far more effective as quantum computers rapidly become more powerful.
6. What’s next?
We have now turned our attention to work that will show that using vectors to represent meaning and requiring compositional interpretability for natural language takes us mathematically natively into the quantum formalism. This does not mean that there doesn't exist an efficient classical method for solving specific tasks, and it may be hard to prove traditional hardness results whenever there is some machine learning involved. This could be something we might have to come to terms with, just as in classical machine learning.
At ԹϺ we possess the most powerful quantum computers currently available. Our recently published roadmap is going to deliver more computationally powerful quantum computers in the short and medium term, as we extend our lead and push towards universal, fault tolerant quantum computers by the end of the decade. We expect to show even better (and larger scale) results when implementing our work on those machines. In short, we foresee a period of rapid innovation as powerful quantum computers that cannot be classically simulated become more readily available. This will likely be disruptive, as more and more use cases, including ones that we might not be currently thinking about, come into play.
Interestingly and intriguingly, we are also pioneering the use of powerful quantum computers in a hybrid system that has been described as a ‘quantum supercomputer’ where quantum computers, HPC and AI work together in an integrated fashion and look forward to using these systems to advance our work in language processing that can help solve the problem with LLM’s that we highlighted at the start of this article.
1 And where do we go next, when we don’t even understand what we are dealing with now? On previous occasions in the history of science and technology, when efficient models without a clear interpretation have been developed, such as the Babylonian lunar theory or Ptolemy’s model of epicycles, these initially highly successful technologies vanished, making way for something else.
2 Note that our conception of compositionality is more general than the usual one adopted in linguistics, which is due to Frege. A discussion can be found in [].
3 For example, using pregroups here as linguistic structure, which are the cups and caps of PQP.
4 That is, using the tensor product of the corresponding vector spaces.
About ԹϺ
ԹϺ, the world’s largest integrated quantum company, pioneers powerful quantum computers and advanced software solutions. ԹϺ’s technology drives breakthroughs in materials discovery, cybersecurity, and next-gen quantum AI. With over 500 employees, including 370+ scientists and engineers, ԹϺ leads the quantum computing revolution across continents.
Blog
|
technical
March 4, 2026
Skinny Logic: Quantum Codes Go on a Diet
In our latest paper, we’ve taken a big step toward large scale fault-tolerant quantum computing, squeezing up to 94 error-detected qubits (and 48 error-corrected qubits) out of just 98 physical qubits, a low-fat encoding that cuts overhead to the bone. With 64 of our logical qubits, we were able to simulate quantum magnetism at a scale that can be exceedingly difficult for classical computers.
The "holy grail" of quantum computing is universal fault-tolerance: the ability to correct errors faster than they occur during any computation. To realize this, we aim to create “logical qubits,” which are groups of entangled physical qubits that share quantum information in a way that protects it. Better protection leads to lower “logical” error rate and greater ability to solve complex problems.
However, it’s never that easy. An unofficial law of physics is “there’s no such thing as a free lunch”. Creating high quality, low error-rate logical qubits often costs many physical qubits, thus reducing the size of calculations you can run, despite your new, lower-than-ever error rates.
With our , we are thrilled to announce that we have hit a key milestone on the ԹϺ roadmap: an ultra-efficient method for creating logical qubits, extracting a whopping 48 error-corrected and 64 error-detected logical qubits out of just 98 physical qubits. Our logical qubits boasted better than “break-even” fidelity, beating their physical counterparts with lower error rates on several different fronts. And still that isn’t the end of the story: we used our 64 error-detected logical qubits in a large-scale quantum magnetism simulation, laying the groundwork for future studies of exotic interactions in materials.
Stacking Wins
To get this world-leading result, we employed a neat trick: ‘nesting’ super efficient quantum error-detecting codes together to make a new, ultra-efficient error-correcting code. Dr. DeCross, a primary author on the paper, said this nesting is like “braiding together ropes made out of ropes made out of ropes”. Physicists call this ‘code concatenation’, and you can think of it as adding layers of protection on top of each other.
To begin, we took the now-famous ‘iceberg code’, a quantum error detection code that gives an almost 1:1 ratio of physical qubits to logical qubits. The iceberg code only detects errors, however, which means that instead of actually correcting errors it lets you throw out bits where errors were detected. To make a code that could both detect and correct errors, we concatenated two iceberg codes together, giving a code that can correct small errors while still boasting a world-record 2:1 physical:logical ratio (physicists call this a “high encoding rate”).
The team then benchmarked the logical qubits, checking large system-scale operations and comparing them to their physical counterparts. This introduces a crucial hurdle to clear: oftentimes, researchers end up with logical qubits that perform *worse* than their physical counterparts. It’s critical that logical qubits actually beat physical ones, after all – that is the whole point!
Thanks to some clever circuit design and our natively high fidelities, the new logical qubits outperformed their physical counterparts in every test we performed, sometimes by a factor of 10 to 100.
Computing Logically
Of course, the whole point is to use our logical qubits for something useful, the ultimate measure of functionality. With 64 error-detected qubits, we performed a simulation of quantum magnetism; a crucial milestone that validates our roadmap.
The team took extra care to perform their simulation in 3 dimensions to best reflect the real-world (often, studies like this will only be in 1D or 2D to make them easier). Problems like this are both incredibly important for expanding our understanding of materials, but are also incredibly hard, as their complexity scales quickly. To make qubits interact as if they are in a 3D material when they are trapped in 2D inside the computer, we used our all-to-all connectivity, a feature that results from our movable qubits.
Maximizing Entanglement
Breaking the encoding rate record and performing a world-leading logical simulation wasn’t enough for the team. For their final feat, the team generated 94 error-detected logical qubits, and entangled them all in a special state called a “GHZ” state (also known as a ‘cat’ state, alluding to Schrödinger’s cat). GHZ states are often used by experts as a simple benchmark for showcasing quantum computing’s unique capacity to use entanglement across many qubits. Our best 94-logical qubit GHZ state boasted a fidelity of 94.9%, crushing its un-encoded counterpart.
Logical Qubits Are the New Normal
Taken together, these results show that we can suppress errors more effectively than ever before, proving that Helios is capable of delivering complex, high-fidelity operations that were previously thought to be years away. While the magnetism simulation was only error-detected, it showcases our ability to protect universal computations with partially fault-tolerant methods. On top of that, the team also demonstrated key error-corrected primitives on Helios at scale.
All of this has real-world implications for the quantum ecosystem: we are working to package these iceberg codes into QCorrect, an upcoming tool that will help developers automatically improve the performance of their own applications.
This is just the beginning: we are officially entering the era of large-scale logical computing. The path to fault-tolerance is no longer just theoretical—it is being built, gate by gate, on Helios.
Hybrid quantum–HPC computing with trapped ions is here
Japan has made bold, strategic investments in both high-performance computing (HPC) and quantum technologies. As these capabilities mature, an important question arises for policymakers and research leaders: how do we move from building advanced machines to demonstrating meaningful, integrated use?
Last year, ԹϺ installed its Reimei quantum computer at a world-class facility in Japan operated by RIKEN, the country’s largest comprehensive research institution. The system was integrated with Japan’s famed supercomputer Fugaku, one of the most powerful in the world, as part of an ambitious national project commissioned by the New Energy and Industrial Technology Development Organization (NEDO), the national research and development entity under the Ministry of Economy, Trade and Industry.
Now, for the first time, a full scientific workflow has been executed across Fugaku, one of the world’s most powerful supercomputers, and Reimei, our trapped-ion quantum computer. This marks a transition from infrastructure development to practical deployment.
Quantum Biology
In this first foray into hybrid HPC-quantum computation, the team explored chemical reactions that occur inside biomolecules such as proteins. Reactions of this type are found throughout biology, from enzyme functions to drug interactions.
Simulating such reactions accurately is extremely challenging. The region where the chemical reaction occurs—the “active site”—requires very high precision, because subtle electronic effects determine the outcome. At the same time, this active site is embedded within a much larger molecular environment that must also be represented, though typically at a lower level of detail.
To address this complexity, computational chemistry has long relied on layered approaches, in which different parts of a system are treated with different methods. In our work, we extended this concept into the hybrid computing era by combining classical supercomputing with quantum computing.
Shifting the Paradigm
While the long-term goal of quantum computing is to outperform classical approaches alone, the purpose of this project was to demonstrate a fully functional hybrid system working as an end-to-end platform for real scientific applications. We believe it is not enough to develop hardware in isolation – we must also build workflows where classical and quantum resources create a whole that is greater than the parts. We believe this is a crucial step for our industry; large-scale national investments in quantum computing must ultimately show how the technology can be embedded within existing research infrastructure.
In this work, the supercomputer Fugaku handled geometry optimization and baseline electronic structure calculations. The quantum computer Reimei was used to enhance the treatment of the most difficult electronic interactions in the active site, those that are known to challenge conventional approximate methods. The entire process was coordinated through ԹϺ’s workflow system , which allows jobs to move efficiently between machines.
Hybrid Computation is Now an Operational Reality
With this infrastructure in place, we are now poised to truly leverage the power of quantum computing. In this instance, the researchers designed the algorithm to specifically exploit the strengths of both the quantum and the classical hardware.
First, the classical computer constructs an approximate description of the molecular system. Then, the quantum computer is used to model the detailed quantum mechanics that the classical computer can’t handle. Together, this improves accuracy, extending the utility of the classical system.
A Path to Hybrid Advantage
Accurate simulation of biomolecular reactions remains one of the major challenges in biochemistry. Although the present study uses simplified systems to focus on methodology, it lays the groundwork for future applications in drug design, enzyme engineering, and photoactive biological systems.
While fully fault-tolerant, large-scale quantum computers are still under development, hybrid approaches allow today’s quantum hardware to augment powerful classical systems, such as Fugaku, to explore meaningful applications. As quantum technology matures, the same workflows can scale accordingly.
High-performance computing centers worldwide are actively exploring how quantum devices might integrate into their ecosystems. By demonstrating coordinated job scheduling, direct hardware access, and workflow orchestration across heterogeneous architectures, this work offers a concrete example of how such integration can be achieved.
As quantum hardware matures, we believe the algorithms and workflows developed here can be extended to increasingly realistic and industrially relevant problems. For Japan’s research ecosystem, this first application milestone signals that hybrid quantum–supercomputing is moving from ambition to implementation.
Automated Quantum Algorithm Discovery for Quantum Chemistry
ܳٳǰ: ԹϺ (alphabetical order): Eric Brunner, Steve Clark, Fabian Finger, Gabriel Greene-Diniz, Pranav Kalidindi, Alexander Koziell-Pipe, David Zsolt Manrique, Konstantinos Meichanetzidis, Frederic Rapp Hiverge (alphabetical order):Alhussein Fawzi, Hamza Fawzi, Kerry He, Bernardino Romera Paredes, Kante Yin
What if every quantum computing researcher had an army of students to help them write efficient quantum algorithms? Large Language Models are starting to serve as such a resource.
ԹϺ’s processors offer world-leading fidelity, and recent experiments show that they have surpassed the limits of classical simulation for certain computational tasks, such as simulating materials. However, access to quantum processors is limited and can be costly. It is therefore of paramount importance to optimise quantum resources and write efficient quantum software. Designing efficient algorithms is a challenging task, especially for quantum algorithms: dealing with superpositions, entanglement, and interference can be counterintuitive.
To this end, our joint team used AI platform for automated algorithm discovery, the Hive, to probe the limits of what can be done in quantum chemistry. The Hive generates optimised algorithms tailored to a given problem, expressed in a familiar programming language, like Python. Thus, the Hive’s outputs allow for increased interpretability, enabling domain experts to potentially learn novel techniques from the AI-discovered solutions. Such AI-assisted workflows lower the barrier of entry for non-domain experts, as an initial sketch of an algorithmic idea suffices to achieve state-of-the-art solutions.
In this initial proof-of-concept study, we demonstrate the advantage of AI-driven algorithmic discovery of efficient quantum heuristics in the context of quantum chemistry, in particular the electronic structure problem. Our early explorations show that the Hive can start from a naïve and simple problem statement and evolve a highly optimised quantum algorithm that solves the problem, reaching chemical precision for a collection of molecules. Our high-level workflow is shown in Figure 1. Specifically, the quantum algorithm generated by the Hive achieves a reduction in the quantum resources required by orders of magnitude compared to current state-of-the-art quantum algorithms. This promising result may enable the implementation of quantum algorithms on near-term hardware that was previously thought impossible due to current resource constraints.
Figure 1: Workflow: A scientist prompts Hiverge's platform, the Hive, with the molecule of interest and a sketch of a quantum algorithm. The goal of the quantum algorithm is to find the ground state energy of the molecule. The Hive evolves the sketch into an efficient version that solves the problem.
The Electronic Structure Problem in Quantum Chemistry
The electronic structure problem is central to quantum chemistry. The goal is to prepare the ground state (the lowest energy state) of a molecule and compute the corresponding energy of that state to chemical precision or beyond. Classically, this is an exponentially hard problem. In particular, classical treatments tend to fall short when there are strong quantum effects in the molecule, and this is where quantum computers may be advantageous.
The paradigm of variational quantum algorithms is motivated by near-term quantum hardware. One starts with a relatively easy-to-prepare initial state. Then, the main part of the algorithm consists of a sequence of parameterised operators representing chemically meaningful actions, such as manipulating electron occupations in the molecular orbitals. These are implemented in terms of parameterised quantum gates. Finally, the energy of the state is measured via the molecule’s energy operator, the “Hamiltonian”, by executing the circuit on a quantum computer and measuring all the qubits on which the circuit is implemented. Taking many measurements, or “shots”, the energy is estimated to the desired precision. The ground state energy is found by iteratively optimising the parameters of the quantum circuit until the energy converges to a minimum value. The general form of such a variational quantum algorithm is illustrated in Figure 2.
Figure 2: A variational quantum algorithm is defined by a function select_next_operator that iteratively constructs a parameterised quantum circuit as a sequence of operators [O1(θ1),O2(θ2),O3(θ3), ...], and a function update_parameters that optimises its parameters; these functions update the quantum circuit and refine it to its final form that prepares the ground state. The Hive evolves sophisticated versions of these functions starting from trivial versions, written in a familiar programming language, producing a novel, efficient variational quantum algorithm that solves the problem.
The main challenge in these frameworks is to design an appropriate quantum circuit architecture, i.e. find an efficient sequence of operators, and an efficient optimisation strategy for its parameters. It is important to minimise the number of quantum operations in any given circuit, as each operation is inherently noisy and the algorithm’s output degrades exponentially. Another important quantum resource to be minimised is the total number of circuits that need to be evaluated to compute the energy values during the optimisation of the circuit parameters, which is time-consuming.
To meet these challenges, we task the Hive with designing a variational quantum algorithm to solve the ground state problem, following the workflow shown in Figure 1. The Hive is a distributed evolutionary process that evolves programs. It uses Large Language Models to generate mutations in the form of edits to an entire codebase. This genetic process selects the fittest programs according to how well they solve a given problem. In our case, the role of the quantum computer is to compute the fitness, i.e., the ground state energy. Importantly, the Hive operates at the level of a programming language; it readily imports and uses all known libraries that a human researcher would use, including ԹϺ’s quantum chemistry platform, InQuanto. In addition, the Hive can accept instructions and requests in natural language, increasing its flexibility. For example, we encouraged it to seek parameter optimisation strategies that avoid estimating gradients, as this incurs significant overhead in terms of circuit evaluations. Intuitively, the interaction between a human scientist and the Hive is analogous to a supervisor and a group of eager and capable students: the supervisor provides guidance at a high level, and the students collaborate and flesh out the general idea to produce a working solution that the supervisor can then inspect.
We find that from an extremely basic starting point, consisting of a skeleton for a variational quantum algorithm, the Hive can autonomously assemble a bespoke variational quantum algorithm, which we call Hive-ADAPT. Specifically, the Hive evolves heuristic functions that construct a circuit as a sequence of quantum operators and optimise its parameters. Remarkably, the Hive converged on a structure resembling the current state-of-the-art, ADAPT-VQE. Crucially, however, Hive-ADAPT substantially outperforms this baseline, delivering significant improvements in chemical precision while reducing quantum resource requirements.
Figure 3: (Top): The measured ground state energy of the molecule in Hartree (Ha) as a function of the bond length in Angstrom (Å), i.e. the length of the O-H and Be-H bonds in H2O and BeH2, respectively. Both ADAPT-VQE and Hive-ADAPT recover the energy curve. (Bottom): The difference between the energy estimated by the quantum algorithms and the reference value computed with the exact FCI method. Hive-ADAPT achieves chemical precision for more bond lengths than ADAPT-VQE (energy below dashed flat lines). Hive-ADAPT was evolved by the Hive to solve a particular set of bond lengths (red circles), and we observe that the same algorithm can also solve the problem on other bond lengths (green circles), showing generalisation over bond lengths.
A molecule’s ground state energy varies with the distances between its atoms, called the “bond length”. For example, for the molecule H2O, the bond length refers to the length of the O-H bond. The Hive was tasked with developing an algorithm for a small set of bond lengths and reaching chemical precision, defined as within 1.6e-3 Hartree (Ha) of the ground state energy computed with the exact Full Configuration Interaction (FCI) algorithm. As we show in Figure 3, remarkably, Hive-ADAPT achieves chemical precision for more bond lengths than ADAPT-VQE. Furthermore, Hive-ADAPT also reaches chemical precision for other “unseen” bond lengths, showcasing the generalisation ability of the evolved quantum algorithm. Our results were obtained from classical simulations of the quantum algorithms, where we used NVIDIA CUDA-Q to leverage the parallelism enabled by GPUs. Further, relative to ADAPT-VQE, Hive-ADAPT exhibits one to two orders of magnitude reduction in quantum resources, such as the number of circuit evaluations and the number of operators used to construct circuits, which is crucial for practical implementations on actual near-term processors.
For molecules such as BeH2 at large Be-H bond lengths, a complex initial state is required for the algorithm to be able to reach the ground state using the available operators. Even in these cases, by leveraging an efficient state preparation scheme implemented in InQuanto, the Hive evolved a dedicated strategy for the preparation of such a complex initial state, given a set of basic operators to achieve the desired chemical precision.
To validate Hive-ADAPT under realistic conditions, we employed ԹϺ’s H2 Emulator, which provides a faithful classical simulator of the H2 quantum computer, characterised by a 1.05e-3 two-qubit gate error rate. Leveraging the Hive's inherent flexibility, we adapted the optimisation strategy to explicitly penalise the number of two-qubit gates—the dominant noise source on near-term hardware—by redefining the fitness function. This constraint guided the Hive to discover a noise-aware algorithm capable of constructing hardware-efficient circuits. We subsequently executed the specific circuit generated by this algorithm for the LiH molecule at a bond length of 1.5 Å with the Partition Measurement Symmetry Verification (PMSV) error mitigation procedure. The resulting energy of -7.8767 ± 0.0031 Ha, obtained using 10,000 shots per circuit with a discard rate below 10% in the PMSV error mitigation procedure, is close to the target FCI energy of -7.8824 Ha and demonstrates the Hive's ability to successfully tailor algorithms that balance theoretical accuracy with the rigorous constraints of hardware noise and approach chemical precision as much as possible with current quantum technology.
For illustration purposes, we show an example of an elaborate code snippet evolved by the Hive starting from a trivial version:
ԹϺ’s in-house quantum chemistry expert, Dr. David Zsolt Manrique, commented,
“I found it amazing that the Hive converged to a domain-expert level idea. By inspecting the code, we see it has identified the well-known perturbative method, ‘MP2’, as a useful guide; not only for setting the initial circuit parameters, but also for ordering excitations efficiently. Further, it systematically and laboriously fine-tuned those MP2-inspired heuristics over many iterations in a way that would be difficult for a human expert to do by hand. It demonstrated an impressive combination of domain expertise and automated machinery that would be useful in exploring novel quantum chemistry methods.”
Looking to the Future
In this initial proof-of-concept collaborative study between ԹϺ and Hiverge, we demonstrate that AI-driven algorithm discovery can generate efficient quantum heuristics. Specifically, we found a great reduction in quantum resources, which is impactful for quantum algorithmic primitives that are frequently reused. Importantly, this approach is highly flexible; it can accommodate the optimisation of any desired quantum resource, from circuit evaluations to the number of operations in a given circuit. This work opens a path toward fully automated pipelines capable of developing problem-specific quantum algorithms optimised for NISQ as well as future hardware.
An important question for further investigation regards transferability and generalisation of a discovered quantum solution to other molecules, going beyond the generalisation over bond lengths of the same molecule that we have already observed. Evidently, this approach can be applied to improving any other near-term quantum algorithm for a range of applications from optimisation to quantum simulation.
We have already demonstrated an error-corrected implementation of quantum phase estimation on quantum hardware, and an AI-driven approach promises further hardware-tailored improvements and optimal use of quantum resources. Beyond NISQ, we envision that AI-assisted algorithm discovery will be a fruitful endeavour in the fault-tolerant regime, as well, where high-level quantum algorithmic primitives (quantum fourier transform, amplitude amplification, quantum signal processing, etc.) are to be combined optimally to achieve computational advantage for certain problems.
Notably, we’ve entered an era where quantum algorithms can be written in high-level programming languages, like ԹϺ’s , and approaches that integrate Large Language Models directly benefit. Automated algorithm discovery is promising for improving routines relevant to the full quantum stack, for example, in low-level quantum control or in quantum error correction.
