Quantinuum researchers tackle AI’s ‘interpretability problem’, helping us build safer systems
The Artificial Intelligence (AI) systems that have recently permeated our lives have a serious problem: they are built in a way that makes them very hard - and sometimes impossible - to understand or interpret. Luckily, our team is tackling this problem, and we’ve just published a new paper that covers the issue in detail.
It turns out that the lack of explainability in machine learning (ML) models, such as ChatGPT or Claude, comes from the way that the systems are built. Their underlying architecture (a neural network) lacks coherent structure. While neural networks can be trained to effectively solve certain tasks, the way they do it is largely (or, from a practical standpoint, almost wholly) inaccessible. This absence of interpretability in modern ML is increasingly a major concern in sensitive areas where accountability is required, such as in finance and the healthcare and pharmaceutical sectors. The “interpretability problem in AI” is therefore a topic of grave worry for large swathes of the corporate and enterprise sector, regulators, lawmakers, and the general public.
These concerns have given birth to the field of eXplainable AI, or XAI, which attempts to solve the interpretability problem through so-called ‘post-hoc’ techniques (where one takes a trained AI model and aims to give explanations for either its overall behavior or individual outputs). This approach, while still evolving, has its own issues due to the approximate nature and fundamental limitations of post-hoc techniques.
The second approach to the interpretability problem is to employ new ML models that are, by design, inherently interpretable from the start. Such an interpretable AI model comes with explicit structure which is meaningful to us “from the outside”. Realizing this in the tech we use every day means completely redesigning how machines learn - creating a new paradigm in AI. As Sean Tull, one of the authors of the paper, stated: “In the best case, such intrinsically interpretable models would no longer even require XAI methods, serving instead as their own explanation, and one of a deeper kind.”
At Quantinuum, we’re continuing work to develop new paradigms in AI while also working to sharpen theoretical and foundational tools that allow us all to assess the interpretability of a given model. In our recent paper, we present a new theoretical framework for both defining AI models and analyzing their interpretability. With this framework, we show how advantageous it is for an AI model to have explicit and meaningful compositional structure.
The idea of composition is explored in a rigorous way using a mathematical approach called “category theory”, which is a language that describes processes and their composition. The category theory approach to interpretability can be accomplished via a graphical calculus which was also developed in part by Quantinuum scientists, and which is finding use cases in everything from gravity to quantum computing.
A fundamental problem in the field of XAI has been that many terms have not been rigorously defined, making it difficult to study - let alone discuss - interpretability in AI. Our paper presents the first known theoretical framework for assessing the compositional interpretability of AI models. With our team’s work, we now have a precise and mathematically defined definition of interpretability that allows us to have these critical conversations.
After developing the framework, our team used it to analyze the full spectrum of ML approaches. We started with Transformers (the “T” in ChatGPT), which are not interpretable – pointing to a serious issue in some of the world’s most widely used ML tools. This is in contrast with (sparse) linear models and decision trees, which we found are indeed inherently interpretable, as they are usually described.
Our team was also able to make precise how other ML models were what they call 'compositionally interpretable'. These include models already studied by our own scientists including DisCo NLP models, causal models, and conceptual space models.
Many of the models discussed in this paper are classical, but more broadly the use of category theory and string diagrams makes these tools very well suited to analyzing quantum models for machine learning. In addition to helping the broader field accurately assess the interpretability of various ML models, the seminal work in this paper will help us to develop systems that are interpretable by design.
This work is part of our broader AI strategy, which includes using AI to improve quantum computing, using quantum computers to improve AI, and – in this case - using the tools of category theory and compositionality to help us better understand AI.
About Quantinuum
Quantinuum, the world’s largest integrated quantum company, pioneers powerful quantum computers and advanced software solutions. Quantinuum’s technology drives breakthroughs in materials discovery, cybersecurity, and next-gen quantum AI. With over 500 employees, including 370+ scientists and engineers, Quantinuum leads the quantum computing revolution across continents.
Typically, Quantum Error Detection (QED) is viewed as a short-term solution—a non-scalable, stop-gap until full fault tolerance is achieved at scale.
That’s just changed, thanks to a serendipitous discovery made by our team. Now, QED can be used in a much wider context than previously thought. Our team made this discovery while studying the contact process, which describes things like how diseases spread or how water permeates porous materials. In particular, our team was studying the quantum contact process (QCP), a problem they had tackled before, which helps physicists understand things like phase transitions. In the process (pun intended), they came across what senior advanced physicist, Eli Chertkov, described as “a surprising result.”
While examining the problem, the team realized that they could convert detected errors due to noisy hardware into random resets, a key part of the QCP, thus avoiding the exponentially costly overhead of post-selection normally expected in QED.
To understand this better, the team developed a new protocol in which the encoded, or logical, quantum circuit adapts to the noise generated by the quantum computer. They quickly realized that this method could be used to explore other classes of random circuits similar to the ones they were already studying.
The team put it all together on System Model H2 to run a complex simulation, and were surprised to find that they were able to achieve near break-even results, where the logically encoded circuit performed as well as its physical analog, thanks to their clever application of QED. Ultimately, this new protocol will allow QED codes to be used in a scalable way, saving considerable computational resources compared to full quantum error correction (QEC).
Researchers at the crossroads of quantum information, quantum simulation, and many-body physics will take interest in this protocol and use it as a springboard for inventing new use cases for QED.
Stay tuned for more, our team always has new tricks up their sleeves.
Learn mode about System Model H2 with this video:
By Konstantinos Meichanetzidis
When will quantum computers outperform classical ones?
This question has hovered over the field for decades, shaping billion-dollar investments and driving scientific debate.
The question has more meaning in context, as the answer depends on the problem at hand. We already have estimates of the quantum computing resources needed for Shor’s algorithm, which has a superpolynomial advantage for integer factoring over the best-known classical methods, threatening cryptographic protocols. Quantum simulation allows one to glean insights into exotic materials and chemical processes that classical machines struggle to capture, especially when strong correlations are present. But even within these examples, estimates change surprisingly often, carving years off expected timelines. And outside these famous cases, the map to quantum advantage is surprisingly hazy.
Researchers at Quantinuum have taken a fresh step toward drawing this map. In a new theoretical framework, Harry Buhrman, Niklas Galke, and Konstantinos Meichanetzidis introduce the concept of “queasy instances” (quantum easy) – problem instances that are comparatively easy for quantum computers but appear difficult for classical ones.
From Problem Classes to Problem Instances
Traditionally, computer scientists classify problems according to their worst-case difficulty. Consider the problem of Boolean satisfiability, or SAT, where one is given a set of variables (each can be assigned a 0 or a 1) and a set of constraints and must decide whether there exists a variable assignment that satisfies all the constraints. SAT is a canonical NP-complete problem, and so in the worst case, both classical and quantum algorithms are expected to perform badly, which means that the runtime scales exponentially with the number of variables. On the other hand, factoring is believed to be easier for quantum computers than for classical ones. But real-world computing doesn’t deal only in worst cases. Some instances of SAT are trivial; others are nightmares. The same is true for optimization problems in finance, chemistry, or logistics. What if quantum computers have an advantage not across all instances, but only for specific “pockets” of hard instances? This could be very valuable, but worst-case analysis is oblivious to this and declares that there is no quantum advantage.
To make that idea precise, the researchers turned to a tool from theoretical computer science: Kolmogorov complexity. This is a way of measuring how “regular” a string of bits is, based on the length of the shortest program that generates it. A simple string like 0000000000 can be described by a tiny program (“print ten zeros”), while the description of a program that generates a random string exhibiting no pattern is as long as the string itself. From there, the notion of instance complexity was developed: instead of asking “how hard is it to describe this string?”, we ask “how hard is it to solve this particular problem instance (represented by a string)?” For a given SAT formula, for example, its polynomial-time instance complexity is the size of the smallest program that runs in polynomial time and decides whether the formula is satisfiable. This smallest program must be consistently answering all other instances, and it is also allowed to declare “I don’t know”.
In their new work, the team extends this idea into the quantum realm by defining polynomial-time quantum instance complexity as the size of the shortest quantum program that solves a given instance and runs on polynomial time. This makes it possible to directly compare quantum and classical effort, in terms of program description length, on the very same problem instance. If the quantum description is significantly shorter than the classical one, that problem instance is one the researchers call “queasy”: quantum-easy and classically hard. These queasy instances are the precise places where quantum computers offer a provable advantage – and one that may be overlooked under a worst-case analysis.
Why “Queasy”?
The playful name captures the imbalance between classical and quantum effort. A queasy instance is one that makes classical algorithms struggle, i.e. their shortest descriptions of efficient programs that decide them are long and unwieldy, while a quantum computer can handle the same instance with a much simpler, faster, and shorter program. In other words, these instances make classical computers “queasy,” while quantum ones solve them efficiently and finding them quantum-easy. The key point of these definitions lies in demonstrating that they yield reasonable results for well-known optimisation problems.
By carefully analysing a mapping from the problem of integer factoring to SAT (which is possible because factoring is inside NP and SAT is NP-complete) the researchers prove that there exist infinitely many queasy SAT instances. SAT is one of the most central and well-studied problems in computer science that finds numerous applications in the real-world. The significant realisation that this theoretical framework highlights is that SAT is not expected to yield a blanket quantum advantage, but within it lie islands of queasiness – special cases where quantum algorithms decisively win.
Algorithmic Utility
Finding a queasy instance is exciting in itself, but there is more to this story. Surprisingly, within the new framework it is demonstrated that when a quantum algorithm solves a queasy instance, it does much more than solve that single case. Because the program that solves it is so compact, the same program can provably solve an exponentially large set of other instances, as well. Interestingly, the size of this set depends exponentially on the queasiness of the instance!
Think of it like discovering a special shortcut through a maze. Once you’ve found the trick, it doesn’t just solve that one path, but reveals a pattern that helps you solve many other similarly built mazes, too (even if not optimally). This property is called algorithmic utility, and it means that queasy instances are not isolated curiosities. Each one can open a doorway to a whole corridor with other doors, behind which quantum advantage might lie.
A North Star for the Field
Queasy instances are more than a mathematical curiosity; this is a new framework that provides a language for quantum advantage. Even though the quantities defined in the paper are theoretical, involving Turing machines and viewing programs as abstract bitstrings, they can be approximated in practice by taking an experimental and engineering approach. This work serves as a foundation for pursuing quantum advantage by targeting problem instances and proving that in principle this can be a fruitful endeavour.
The researchers see a parallel with the rise of machine learning. The idea of neural networks existed for decades along with small scale analogue and digital implementations, but only when GPUs enabled large-scale trial and error did they explode into practical use. Quantum computing, they suggest, is on the cusp of its own heuristic era. “Quristics” will be prominent in finding queasy instances, which have the right structure so that classical methods struggle but quantum algorithms can exploit, to eventually arrive at solutions to typical real-world problems. After all, quantum computing is well-suited for small-data big-compute problems, and our framework employs the concepts to quantify that; instance complexity captures both their size and the amount of compute required to solve them.
Most importantly, queasy instances shift the conversation. Instead of asking the broad question of when quantum computers will surpass classical ones, we can now rigorously ask where they do. The queasy framework provides a language and a compass for navigating the rugged and jagged computational landscape, pointing researchers, engineers, and industries toward quantum advantage.
From September 16th – 18th, Quantum World Congress (QWC) brought together visionaries, policymakers, researchers, investors, and students from across the globe to discuss the future of quantum computing in Tysons, Virginia.
Quantinuum is forging the path to universal, fully fault-tolerant quantum computing with our integrated full-stack. With our quantum experts were on site, we showcased the latest on Quantinuum 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.
Highlights from QWC
Dr. Patty Lee Named the Industry Pioneer in Quantum
The Quantum Leadership Awards celebrate visionaries transforming quantum science into global impact. This year at QWC, Dr. Patty Lee, our Chief Scientist for Hardware Technology Development, was named the Industry Pioneer in Quantum! This honor celebrates her more than two decades of leadership in quantum computing and her pivotal role advancing the world’s leading trapped-ion systems. Watch the Award Ceremony here.
Keynote with Quantinuum's CEO, Dr. Rajeeb Hazra
At QWC 2024, Quantinuum’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, Raj shared the progress we’ve made over the last year in advancing quantum computing on both commercial and technical fronts and exciting insights on what’s to come from Quantinuum. Access the full session here.
Panel Session: Policy Priorities for Responsible Quantum and AI
As part of the Track Sessions on Government & Security, Quantinuum’s Director of Government Relations, Ryan McKenney, discussed “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
During the Track Session on Industry Advancement, Quantinuum’s Chief Legal Officer, Kaniah Konkoly-Thege, and Director of Government Relations, Ryan McKenney, discussed the importance of “Establishing a Pro-Innovation Regulatory Framework”.