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 , 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 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 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 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
, the world’s largest integrated quantum company, pioneers powerful quantum computers and advanced software solutions. ’s technology drives breakthroughs in materials discovery, cybersecurity, and next-gen quantum AI. With over 500 employees, including 370+ scientists and engineers, leads the quantum computing revolution across continents.
Blog
September 15, 2025
Quantum World Congress 2025
From September 16th – 18th, (QWC) will bring together visionaries, policymakers, researchers, investors, and students from across the globe to discuss the future of quantum computing in Tysons, Virginia.
is forging the path to universal, fully fault-tolerant quantum computing with our integrated full-stack. Join our quantum experts for the below sessions and at Booth #27 to discuss the latest on Systems, the world’s highest-performing, commercially available quantum computers, our new software stack featuring the key additions of Guppy and Selene, our path to error correction, and more.
Wednesday, September 17th
Keynote with 's CEO, Dr. Rajeeb Hazra 9:00 – 9:20am ET | Main Stage
At QWC 2024, ’s President & CEO, Dr. Rajeeb “Raj” Hazra, took the stage to showcase our commitment to advancing quantum technologies through the unveiling of our roadmap to universal, fully fault-tolerant quantum computing by the end of this decade. This year at QWC 2025, join Raj on the main stage to discover the progress we’ve made over the last year in advancing quantum computing on both commercial and technical fronts and be the first to hear exciting insights on what’s to come from .
Panel Session: Policy Priorities for Responsible Quantum and AI 1:00 – 1:30pm ET | Maplewood Hall
As part of the Track Sessions on Government & Security, ’s Director of Government Relations, Ryan McKenney, will discuss “Policy Priorities for Responsible Quantum and AI” with Jim Cook from Actions to Impact Strategies and Paul Stimers from Quantum Industry Coalition.
Fireside Chat: Establishing a Pro-Innovation Regulatory Framework 4:00 – 4:30pm ET | Vault Theater
During the Track Session on Industry Advancement, ’s Chief Legal Officer, Kaniah Konkoly-Thege, and Director of Government Relations, Ryan McKenney, will take the stage to discuss the importance of “Establishing a Pro-Innovation Regulatory Framework”.
In the world of physics, ideas can lie dormant for decades before revealing their true power. What begins as a quiet paper in an academic journal can eventually reshape our understanding of the universe itself.
In 1993, nestled deep in the halls of Yale University, physicist Subir Sachdev and his graduate student Jinwu Ye stumbled upon such an idea. Their work, originally aimed at unraveling the mysteries of “spin fluids”, would go on to ignite one of the most surprising and profound connections in modern physics—a bridge between the strange behavior of quantum materials and the warped spacetime of black holes.
Two decades after the paper was published, it would be pulled into the orbit of a radically different domain: quantum gravity. Thanks to work by renowned physicist Alexei Kitaev in 2015, the model found new life as a testing ground for the mind-bending theory of holography—the idea that the universe we live in might be a projection, from a lower-dimensional reality.
Holography is an exotic approach to understanding reality where scientists use holograms to describe higher dimensional systems in one less dimension. So, if our world is 3+1 dimensional (3 spatial directions plus time), there exists a 2+1, or 3-dimensional description of it. In the words of Leonard Susskind, a pioneer in quantum holography, "the three-dimensional world of ordinary experience—the universe filled with galaxies, stars, planets, houses, boulders, and people—is a hologram, an image of reality coded on a distant two-dimensional surface."
The “SYK” model, as it is known today, is now considered a quintessential framework for studying strongly correlated quantum phenomena, which occur in everything from superconductors to strange metals—and even in black holes. In fact, The SYK model has also been used to study one of physics’ true final frontiers, quantum gravity, with the authors of the paper calling it “a paradigmatic model for quantum gravity in the lab.”
The SYK model involves Majorana fermions, a type of particle that is its own antiparticle. A key feature of the model is that these fermions are all-to-all connected, leading to strong correlations. This connectivity makes the model particularly challenging to simulate on classical computers, where such correlations are difficult to capture. Our quantum computers, however, natively support all-to-all connectivity making them a natural fit for studying the SYK model.
Now, 10 years after Kitaev’s watershed lectures, we’ve made new progress in studying the SYK model. In a new paper, . By exploiting our system’s native high fidelity and all-to-all connectivity, as well as our scientific team’s deep expertise across many disciplines, we were able to study the SYK model at a scale three times larger than the previous best experimental attempt.
While this work does not exceed classical techniques, it is very close to the classical state-of-the-art. The biggest ever classical study was done on 64 fermions, while our recent result, run on our smallest processor (System Model H1), included 24 fermions. Modelling 24 fermions costs us only 12 qubits (plus one ancilla) making it clear that we can quickly scale these studies: our System Model H2 supports 56 qubits (or ~100 fermions), and Helios, which is coming online this year, will have over 90 qubits (or ~180 fermions).
However, working with the SYK model takes more than just qubits. The SYK model has a complex Hamiltonian that is difficult to work with when encoded on a computer—quantum or classical. Studying the real-time dynamics of the SYK model means first representing the initial state on the qubits, then evolving it properly in time according to an intricate set of rules that determine the outcome. This means deep circuits (many circuit operations), which demand very high fidelity, or else an error will occur before the computation finishes.
Our cross-disciplinary team worked to ensure that we could pull off such a large simulation on a relatively small quantum processor, laying the groundwork for quantum advantage in this field.
First, the team adopted a to run the simulation. By using random sampling, among other methods, the TETRIS algorithm allows one to compute the time evolution of a system without the pernicious discretization errors or sizable overheads that plague other approaches. TETRIS is particularly suited to simulating the SYK model because with a high level of disorder in the material, simulating the SYK Hamiltonian means averaging over many random Hamiltonians. With TETRIS, one generates random circuits to compute evolution (even with a deterministic Hamiltonian). Therefore, when applying TETRIS on SYK, for every shot one can just generate a random instance of the Hamiltonain, and generate a random circuit on TETRIS at the same time. This simple approach enables less gate counts required per shot, meaning users can run more shots, naturally mitigating noise.
In addition, the team “sparsified” the SYK model, which means “pruning” the fermion interactions to reduce the complexity while still maintaining its crucial features. By combining sparsification and the TETRIS algorithm, the team was able to significantly reduce the circuit complexity, allowing it to be run on our machine with high fidelity.
They didn’t stop there. The team also proposed two new noise mitigation techniques, ensuring that they could run circuits deep enough without devolving entirely into noise. The two techniques both worked quite well, and the team was able to show that their algorithm, combined with the noise mitigation, performed significantly better and delivered more accurate results. The perfect agreement between the circuit results and the true theoretical results is a remarkable feat coming from a co-design effort between algorithms and hardware.
As we scale to larger systems, we come closer than ever to realizing quantum gravity in the lab, and thus, answering some of science’s biggest questions.
At , we pay attention to every detail. From quantum gates to teleportation, we work hard every day to ensure our quantum computers operate as effectively as possible. This means not only building the most advanced hardware and software, but that we constantly innovate new ways to make the most of our systems.
A key step in any computation is preparing the initial state of the qubits. Like lining up dominoes, you first need a special setup to get meaningful results. This process, known as state preparation or “state prep,” is an open field of research that can mean the difference between realizing the next breakthrough or falling short. Done ineffectively, state prep can carry steep computational costs, scaling exponentially with the qubit number.
Recently, our algorithm teams have been tackling this challenge from all angles. We’ve published three new papers on state prep, covering state prep for chemistry, materials, and fault tolerance.
In the , our team tackled the issue of preparing states for quantum chemistry. Representing chemical systems on gate-based quantum computers is a tricky task; partly because you often want to prepare multiconfigurational states, which are very complex. Preparing states like this can cost a lot of resources, so our team worked to ensure we can do it without breaking the (quantum) bank.
To do this, our team investigated two different state prep methods. The first method uses , implemented to save computational costs. The second method exploits the sparsity of the molecular wavefunction to maximize efficiency.
Once the team perfected the two methods, they implemented them in InQuanto to explore the benefits across a range of applications, including calculating the ground and excited states of a strongly correlated molecule (twisted C_2 H_4). The results showed that the “sparse state preparation” scheme performed especially well, requiring fewer gates and shorter runtimes than alternative methods.
In the , our team focused on state prep for materials simulation. Generally, it’s much easier for computers to simulate materials that are at zero temperature, which is, obviously, unrealistic. Much more relevant to most scientists is what happens when a material is not at zero temperature. In this case, you have two options: when the material is steadily at a given temperature, which scientists call thermal equilibrium, or when the material is going through some change, also known as out of equilibrium. Both are much harder for classical computers to work with.
In this paper, our team looked to solve an outstanding problem: there is no standard protocol for preparing thermal states. In this work, our team only targeted equilibrium states but, interestingly, they used an out of equilibrium protocol to do the work. By slowly and gently evolving from a simple state that we know how to prepare, they were able to prepare the desired thermal states in a way that was remarkably insensitive to noise.
Ultimately, this work could prove crucial for studying materials like superconductors. After all, no practical superconductor will ever be used at zero temperature. In fact, we want to use them at room temperature – and approaches like this are what will allow us to perform the necessary studies to one day get us there.
Finally, as we advance toward the fault-tolerant era, we encounter a new set of challenges: making computations fault-tolerant at every step can be an expensive venture, eating up qubits and gates. In the , our team made fault-tolerant state preparation—the critical first step in any fault-tolerant algorithm—roughly twice as efficient. With our new “flag at origin” technique, gate counts are significantly reduced, bringing fault-tolerant computation closer to an everyday reality.
The method our researchers developed is highly modular: in the past, to perform optimized state prep like this, developers needed to solve one big expensive optimization problem. In this new work, we’ve figured out how to break the problem up into smaller pieces, in the sense that one now needs to solve a set of much smaller problems. This means that now, for the first time, developers can prepare fault-tolerant states for much larger error correction codes, a crucial step forward in the early-fault-tolerant era.
On top of this, our new method is highly general: it applies to almost any QEC code one can imagine. Normally, fault-tolerant state prep techniques must be anchored to a single code (or a family of codes), making it so that when you want to use a different code, you need a new state prep method. Now, thanks to our team’s work, developers have a single, general-purpose, fault-tolerant state prep method that can be widely applied and ported between different error correction codes. Like the modularity, this is a huge advance for the whole ecosystem—and is quite timely given our recent advances into true fault-tolerance.
This generality isn’t just applicable to different codes, it’s also applicable to the states that you are preparing: while other methods are optimized for preparing only the |0> state, this method is useful for a wide variety of states that are needed to set up a fault tolerant computation. This “state diversity” is especially valuable when working with the best codes – codes that give you many logical qubits per physical qubit. This new approach to fault-tolerant state prep will likely be the method used for fault-tolerant computations across the industry, and if not, it will inform new approaches moving forward.
From the initial state preparation to the final readout, we are ensuring that not only is our hardware the best, but that every single operation is as close to perfect as we can get it.
