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 , 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
|
partnership
November 17, 2025
Powering Hybrid Quantum AI Supercomputing with NVIDIA
is focusing on redefining what’s possible in hybrid quantum–classical computing by integrating ’s best-in-class systems with high-performance NVIDIA accelerated computing to create powerful new architectures that can solve the world’s most pressing challenges.
The launch of Helios, Powered by Honeywell, the world’s most accurate quantum computer, marks a major milestone in quantum computing. Helios is now available to all customers through the cloud or on-premise deployment, launched with a go-to-market offering that seamlessly pairs Helios with the , targeting specific end markets such as drug discovery, finance, materials science, and advanced AI research.
We are also working with NVIDIA to adopt , an open system architecture, as a standard for advancing hybrid quantum-classical supercomputing. Using this technology with Guppy and the , has implemented NVIDIA accelerated computing across Helios and future systems to perform real-time decoding for quantum error correction.
In an industry-first demonstration, an NVIDIA GPU-based decoder integrated in the Helios control engine improved the logical fidelity of quantum operations by more than 3% — a notable gain given Helios’ already exceptionally low error rate. These results demonstrate how integration with NVIDIA accelerated computing through NVQLink can directly enhance the accuracy and scalability of quantum computation.
This unique collaboration spans the full technology stack. ’s next-generation software development environment allows users to interleave quantum and GPU-accelerated classical computations in a single workflow. Developers can build hybrid applications using tools such as NVIDIA CUDA-Q, , and ’s Guppy, to make advanced quantum programming accessible to a broad community of innovators.
The collaboration also reaches into applied research through the (NVAQC), where an NVIDIA GB200 NVL72 supercomputer can be paired with ’s Helios to further drive hybrid quantum-GPU research, including the development of breakthrough quantum-enhanced AI applications.
A recent achievement illustrates this potential: The ADAPT-GQE framework, a transformer-based Generative Quantum AI (GenQAI) approach, uses a Generative AI model to efficiently synthesize circuits to prepare the ground state of a chemical system on a quantum computer. Developed by , NVIDIA, and a pharmaceutical industry leader—and leveraging NVIDIA CUDA-Q with GPU-accelerated methods—ADAPT-GQE achieved a 234x speed-up in generating training data for complex molecules. The team used the framework to explore imipramine, a molecule crucial to pharmaceutical development. The transformer was trained on imipramine conformers to synthesize ground state circuits at orders of magnitude faster than ADAPT-VQE, and the circuit produced by the transformer was run on Helios to prepare the ground state using InQuanto, 's computational chemistry platform.
From collaborating on hardware and software integrations to GenQAI applications, the collaboration between and NVIDIA is building the bridge between classical and quantum computing and creating a future where AI becomes more expansive through quantum computing, and quantum computing becomes more powerful through AI.
The earliest works on quantum error correction showed that by combining many noisy physical qubits into a complex entangled state called a "logical qubit," this state could survive for arbitrarily long times. QEC researchers devote much effort to hunt for codes that function well as "quantum memories," as they are called. Many promising code families have been found, but this is only half of the story.
Being able to keep a qubit around for a long time is one thing, but to realize the theoretical advantages of quantum computing we need to run quantum circuits. And to make sure noise doesn't ruin our computation, these circuits need to be run on the logical qubits of our code. This is often much more challenging than performing gates on the physical qubits of our device, as these "logical gates" often require many physical operations in their implementation. What's more, it often is not immediately obvious which logical gates a code has, and so converting a physical circuit into a logical circuit can be rather difficult.
Some codes, like the famous , are good quantum memories and also have easy logical gates. The drawback is that the ratio of physical qubits to logical qubits (the "encoding rate") is low, and so many physical qubits are required to implement large logical algorithms. High-rate codes that are good quantum memories have also been found, but computing on them is much more difficult. The holy grail of QEC, so to speak, would be a high-rate code that is a good quantum memory and also has easy logical gates. Here, we make progress on that front by developing a new code with those properties.
Building on prior error correcting codes
A recent work from QEC researchers introduced . The underlying construction method for these codes, called the "symplectic double cover," also provided a way to obtain logical gates that are well suited for 's QCCD architecture. Namely, these "SWAP-transversal" gates are performed by applying single qubit operations and relabeling the physical qubits of the device. Thanks to the all-to-all connectivity facilitated through qubit movement on the QCCD architecture, this relabeling can be done in software essentially for free. Combined with extremely high fidelity (~1.2 x10-5) single-qubit operations, the resulting logical gates are similarly high fidelity.
Given the promise of these codes, we take them a step further in our . We combine the symplectic double codes with the [[4,2,2]] Iceberg code using a procedure called "code concatenation". A concatenated code is a bit like nesting dolls, with an outer code containing codes within it---with these too potentially containing codes. More technically, in a concatenated code the logical qubits of one code act as the physical qubits of another code.
The new codes, which we call "concatenated symplectic double codes", were designed in such a way that they have many of these easily-implementable SWAP-transversal gates. Central to its construction, we show how the concatenation method allows us to "upgrade" logical gates in terms of their ease of implementation; this procedure may provide insights for constructing other codes with convenient logical gates. Notably, the SWAP-transversal gate set on this code is so powerful that only two additional operations (logical T and S) are necessary for universal computation. Furthermore, these codes have many logical qubits, and we also present numerical evidence to suggest that they are good quantum memories.
Concatenated symplectic double codes have one of the easiest logical computation schemes, and we didn’t have to sacrifice rate to achieve it. Looking forward in our roadmap, we are targeting hundreds of logical qubits at ~ 1x 10-8 logical error rate by 2029. These codes put us in a prime position to leverage the best characteristics of our hardware and create a device that can achieve real commercial advantage.
at SC25: Advancing the Integration of Quantum and High-Performance Computing
Every year, the brings together the global supercomputing community to explore the technologies driving the future of computing.
Join at this year’s conference, taking place November 16th – 21st in St. Louis, Missouri, where we will showcase how our quantum hardware, software, and partnerships are helping define the next era of high-performance and quantum computing.
Visit in the Expo Hall
The team will be on-site at booth #4432 to showcase how we’re building the bridge between HPC and quantum.
Live demo unit of our quantum hardware
Our new Helios replica, providing an up-close look at the design behind our next-generation system
The Helios chip, highlighting the innovation driving the world’s most advanced trapped-ion quantum computers
On Tuesday and Wednesday, our quantum computing experts will host daily tutorials at our booth on Helios, our next-generation hardware platform, Nexus, our all-in-one quantum computing platform, and Hybrid Workflows, featuring the integration of NVIDIA CUDA-Q with Systems.
Join our team as they share insights on the opportunities and challenges of quantum integration within the HPC ecosystem:
Panel Session: The Quantum Era of HPC: Roadmaps, Challenges and Opportunities in Navigating the Integration Frontier November 19th | 10:30 – 12:00pm CST
During this , Kentaro Yamamoto from , will join experts from Lawrence Berkeley National Laboratory, IBM, QuEra, RIKEN, and Pawsey Supercomputing Research Centre to explore how quantum and classical systems are being brought together to accelerate scientific discovery and industrial innovation.
BoF Session: Bridging the Gap: Making Quantum-Classical Hybridization Work in HPC November 19th | 5:15 – 6:45pm CST
Quantum-classical hybrid computing is moving from theory to reality, yet no clear roadmap exists for how best to integrate quantum processing units (QPUs) into established HPC environments. In this , co-led by ’s Grahame Vittorini and representatives from BCS, DOE, EPCC, Inria, ORNL NVIDIA, and RIKEN we hope to bring together a global community of HPC practitioners, system architects, quantum computing specialists and workflow researchers, including participants in the Workflow Community Initiative, to assess the state of hybrid integration and identify practical steps toward scalable, impactful deployment.
