QMUL Centre for Fundamentals of AI and Computational Theory News

Vasileios Klimis presents paper on "Beyond Specification Conformance" in Rio de Janeiro

Thu, 16 Apr 2026 23:00:00 +0100

Vasileios Klimis has presented a paper on "Beyond Specification Conformance" at the 48th IEEE/ACM International Conference on Software Engineering in Rio de Janeiro It introduces a new way, based on a mathematical logic, that complements specifications to take into account the needs of users to improve the quality of software. Abstract We build software to serve people. Yet time and again, systems pass every test, meet every requirement, and still fail the people who use them. Not because of bugs. Not because of careless engineering. But because the software did exactly what it was told -- and what it was told was never the whole story. This work asks a simple but surprisingly hard question: how do we formally measure the gap between what a system does and what people actually expected it to do? I introduce Semantic Expectation Logic (SEL) -- a framework that treats stakeholder expectations as a first-class, formally reasoned artifact. Not buried in a wiki. Not assumed in a style guide. Not whispered in a standup. Formally captured, systematically elicited, and quantifiably tested. SEL does not replace specifications. It addresses what specifications were never designed to hold - the conventions, assumptions, and shared understanding that developers, users, and operators carry in their heads, and that no test suite has ever been written to check. The result is a new way of thinking about software quality. Not just: did we build the system right? But: did we build the right system -- for the people who have to live with it? Read the Computer Science for Fun (CS4FN) article on the paper. Read the paper itself. Reference Klimis V. Beyond Spec Conformance: A Logic for Validating Stakeholder Expectations. In IEEE/ACM 48th International Conference on Software Engineering (ICSE-NIER), Rio de Janeiro, Brazil, 2026.

1st AI: Brains and Bits Symposium

Sun, 12 Apr 2026 23:00:00 +0100

The first QMUL AI: Brains and Bits symposium was held on 13 April 2026. It brought together researchers from across Science and Engineering to discuss the Fundamentals of AI including issues of how AI works and how it should work. It included talks from investigators from Biology, the Blizzard Institute, Physics, Maths, Electronic Engineering and Computer Science as well as whole and small group discussion sessions, so brought together wide-ranging disciplinary viewpoints to the fundamentals of AI. Topics covered ranged from animal based neuroscience and the fundamental principles of human learning to collective AI and the social behaviour of machines. Abhishek Banerjee first spoke about the neuroscience behind active learning in mammals, driven by life challenges, resources and social interaction. Research on, for example, rats allows parts of the brain to be switched off to help us deeply understand how mammalian brains work and so how AI could. Now computer science models can help better understand this, and how this might lead to better understandding of future agentic AI. Early models were based on feed-forward mechanisms from sensory input to flexible bahaviour. Now models are bidirectional and this happens at the cell neuron level as well as higher levels. Iran Roman, by contrast, spoke about models of machine learning that learn dynamically. Modern machine learning is based on back propagation but older models were more sophisticated. Rather than having a passive learning phase followed by an action phase when the learning is used, in dynamic models learning continues to happen so is reactive. Results show this can be as effective as current algorithms and has potential to be more flexible approach for the future as it is interprettable controllable and optimizable. Andrea Benucci then discussed the way the human brain works from a psychology point of view and how the brain processes visual perception experiences through two separate streams, one focussing on the 'what' and another on the 'why' that follow different pathways in the brain. Our brains also take input about both bodily motion and eye motion (top-down motor-signals) to maintain a stable perception of what is being seen (a bottom-up visualise signal), rather than just working from a visual signal. This understanding has applications in embodied agent systems such as robotic and self-driving car systems, giving ideas for new architectures so new ways for AI to work. Vito Latora finished the first session with a network science take on the fundamentals of AI talking about how there is limited work on collective intelligence and using the wisdom of the crowds. This linked to later points made by David Berman, of aiming for social intelligence rather than ever aiming for "bigger brains" in AI. The former is likely to be the long term way forward, so more research is needed understanding the fundamentals such as how behaviour spreads, what collective AI will look like and how might human-AI collective intelligence might work. In the afternoon, Chris White first spoke about the ongoing importance of meta-science and how AI was able to contribute. What is needed, for example, to understand public attitudes to science is to understand people's stances on issues from social media posts and the like. What is generally done is instead sentiment analysis which is consistently bad at predicting stance. David Berman outlined his personal journey moving from studying string theory to working in industry applying AI and physics-informed approaches, including for formal mathematics. He raised the issue of understanding the scaling properties of systems: do we scale communication speeds, architecture or number of people connected. For example, if we double the speed of communicating do we double the complexity of the emergent society? The rates that AI is scaling currently are essentially insane - doubling every 7 months. Projects need to take the rapid changes into account. Boris Khoruzhenko then discussed modelling of complex systems and how interactions can be replaced by random matrices or random functions to investigate them, answering questions such as whether a large system will be stable eg will small disturbances only lead to small changes. Finally, before the participants split into small groups to discuss issues that emerged, Adrian Baule talked about the application of statistical mechanics of non-equilibrium systems and how behaviour emerges from the behaviour of the components. AI models can potentially help understand such systems, as well as it being something that needs to be understood about the properties that emerge of network based and social AI architectures. The separate talks led to a great deal of interdisiciplinary discussion about the issues arising with potential for a range of future collaborations.

Raymond Hu awarded a grant to investigate Distributed Dynamic Software Updates worth £646,000 ...

Mon, 30 Mar 2026 23:00:00 +0100

Raymond Hu has been awarded a New Investigator research grant worth £646,000 by EPSRC. The project is called "DymSUM - Distributed Dynamic Software Updates using Multiparty Session Types" and investigates a theory-based approach to how distributed software can evolve while running. The project involves collaboration with industrial partners at Monzo Bank and SAP. Distributed systems are at the heart of our infrastructures and society, encompassing, for example, the many Internet and mobile applications that we rely on in our daily lives. A crucial characteristic of modern Distributed Systems is dynamic evolution: many important Distributed Systems are designed to evolve - while the system remains running - in both their program source code (e.g., updates to add or improve features and fix bugs) and execution configuration (e.g., migration of communication links and processes). The project focuses on what are known as Multiparty session types (MSTs). They are a type systems approach to message passing programming. This offers a theoretical grounding for formalising communication protocols and verifying that distributed programs are protocol-compliant. This project will investigate the development of MST-based techniques and tools for formal specification and safe implementation of dynamically-evolving Distributed Sysstems (DDS). The project will support a PDRA and a PhD for three years, starting later this year.

Søren Riis and Marc Roth are joint winners of "Humanity's Last Exam" Competition

Fri, 20 Mar 2026 00:00:00 +0100

Søren Riis and Marc Roth are both joint winners of the SafeAI Benchmark Competition "Humanity's Last Exam". The linked paper "A benchmark of expert-level academic questions to assess AI capabilities", co-authored by FACT researchers Søren Riis and Marc Roth, has been published in Nature. This paper accompanies the benchmark set created in Humanity's Last Exam (HLE), an initiative designed to push artificial intelligence to its limits by challenging it with expert-level questions. AI systems are typically evaluated based on benchmark questions that assess their intelligence and performance. However, as AI models have rapidly advanced, existing benchmarks have become too easy. The HLE competition aimed to change this by curating a new benchmark set of exceptionally difficult questions. The competition attracted more than 1,000 researchers and experts, who submitted questions spanning over 100 subjects. The selection process involved three stages: 1. AI Evaluation: five of the best AI models (late 2024) attempted each question. If all failed, the question advanced. 2. Expert Review: experts refined and assessed the questions and answers. 3. Final Selection: a panel of experts and organisers made the final call. Out of over 70,000 submitted questions to stage 1, only 2,500 made it into the final benchmark, with the top 50 declared as winners, each earning a prize. All contributors were invited to join the paper accompanying the competition as co-authors. Søren and Marc were the sole participants from QMUL. Both contributed multiple questions, and both are joint winners of the competition. Moreover, one of Marc Roth's questions has further been selected to be featured in the Nature paper. At the time the first version of the benchmark set had been finalised (early 2025), the best performing AIs were Open AI o1 and Deepseek R1 which answered, respectively, 8% and 8.5% of the questions correctly. One year later, Gemini 3 Pro achieved a staggering 38.3%. In fact, the true performance might be even better since the benchmark set might still contain a small fraction of ambiguous questions and questions where the given expert answers are partially incomplete or incorrect, mainly in the areas of text-only chemistry and biology questions. The HLE team has therefore transitioned to a dynamic rolling basis for quality control and improvement over the coming years. Read more about it in our article on the competition for the general public.

The effect of images on the ability of Large Language Models to predict human judgements of how ...

Thu, 19 Mar 2026 00:00:00 +0100

A new paper, "Predicting Sentence Acceptability Judgments in Multimodal Contexts", by a team including Shalom Lappin from the Centre for Fundamental AI and Computational Theory, explores the effect of images on the ability of Large Language Models (LLMs) to predict the ratings humans give to sentences over how acceptable they are. For example, how acceptable are each of these sentences (given an image to give context)? "But the answer seems to be no: Reeves lets it be known she requests no costings on raising the three forbidden taxes." (original sentence) "But the reply does not seem: reeves suggests that it would not cost to raise the three banned taxes." (modified sentence) The team found that unlike when a written context is given, images that provide context had little if any impact on the ratings given by humans of how acceptable the sentences were. Different kinds of LLMs were able to predict human acceptability judgments very accurately. However, in general, their performance was slightly better when images are removed! Moreover, the distribution of LLM judgments varies among models. The LLM Qwen resembled human patterns, but others diverged from them. This experimental work suggests that a larger gap exists between the internal representations of LLMs and their generated predictions when images are present to give visual contexts. It suggests interesting points of similarity and of difference between the way humans and LLMs process sentences in multimodal contexts. Read the paper "Predicting Sentence Acceptability Judgments in Multimodal Contexts" on arXiv

Søren Riis and Bei Zhou show AI's game-playing still has flaws

Fri, 13 Mar 2026 00:00:00 +0100

New research published in Machine Learning shows pattern learning is not enough to train AI to tackle games – and abstract representations or hybrid approaches may help. See full story at: https://www.qmul.ac.uk/media/news/2026/science-and-engineering/se/ais-game-playing-still-has-flaws-research-shows.html

A new efficient algorithm for counting network motifs

Sat, 31 Jan 2026 00:00:00 +0100

A paper about a new algorithm that counts important patterns in networks, by a team including Marc Roth from the Centre has been accepted by SODA 2026, the Symposium on Discrete Algorithms (the premier venue for foundations of algorithms research). "Network Motifs" are patterns in the structure of networks. A simple example of a motif is where two points are linked by a series of single hops, but where there is also a direct shortcut link that takes you there in one hop. If this pattern appears more often than expected in a network then it would be a motif of that network. Motifs are important in real-life networks such as social networks or genetic networks (that control how molecules interact with each other inside cells). Network motifs appear more often in such real-life networks than in random networks. The number of times certain motifs appear in a network correlates with a variety of important properties of that network as a whole, such as the ways nodes cluster together in communication networks, or how certain types of cancer are more likely in metabolic networks. Consequently, the computational problem of computing the number of times given motifs appears in large networks has received a lot of attention over the last 20 years. So far, the state of the art applies almost exclusively to networks in which data can be represented as a graph, that is, via individual datapoints that are organised in binary connections where edges of the graph link only two nodes. In this new paper, a novel algorithm has been developed that operates on "hypergraphs", that is, on data organised in a way where one edge can link between more than two nodes, such as in relational databases. Moreover, the authors have proved that the newly developed algorithm is optimal with respect to its running time under standard assumptions from complexity theory, related to the P vs NP conjecture. There can be no faster algorithm that does the same thing. The paper "The Parameterised Complexity of Counting Small Sub-Hypergraphs" can be found at: https://doi.org/10.1137/1.9781611978971.72

Raymond Hu's research on the design and implementation of the Go programming language published ...

Fri, 16 Jan 2026 00:00:00 +0100

Raymond Hu and his coauthors presented their work, that is part of an ongoing collaboration with industry, on "Welterweight Go: Boxing, Structural Subtyping, and Generics" at the 53rd ACM SIGPLAN Symposium on Principles of Programming Languages (POPL 2026) in Rennes, France. POPL is a top A* conference related to the theory, practice and implementation of programming languages. This work is part of an ongoing collaboration with Google on the design and implementation of the Go programming language that is widely used in industry. It builds on their prior work that had the impact of shaping the design of generics for Go first released in Go 1.18. In this paper, the team first develop WG, a formal model of Go expanded to include recently implemented features such as generic type unions and type sets. They also further extend it to show how popularly requested features such as generic methods can be safely incorporated. They then introduce LWG, a lower level formal model of the Go runtime and its mechanisms for managing coercions between interface types and struct types that are key to Go. Finally, they develop a compilation strategy from WG to LWG that supports the proposed extensions and lifts the expressiveness restrictions imposed by the existing monomorphisation approach. Reference Raymond Hu, Julien Lange, Bernardo Toninho, Philip Wadler, Robert Griesemer, Keith Randall (2026) Welterweight Go: Boxing, Structural Subtyping, and Generics. In Proceedings of the ACM on Programming Languages, Volume 10, Issue POPL, Article No.: 79, Pages 2295 - 2322. https://doi.org/10.1145/3776721

PhD student Ben Hayes wins best student paper award

Sun, 05 Oct 2025 23:00:00 +0100

A paper co-authored by AI and Music (AIM) Centre for Doctoral Training PhD student Ben Hayes with supervisors George Fazekas and Charis Saitis received the best student paper award at the 26th International Society for Music Information Retrieval Conference, ISMIR 2025 in Daejeon, Korea. The conference is a flagship conference in the field of Music Informatics. The paper addresses the challenging and ill-posed problem of estimating audio synthesiser parameters given sound examples, where multiple parameter configurations can produce identical sounds due to intrinsic symmetries in synthesiser design. By explicitly modelling these symmetries, particularly permutation invariance across repeated components like oscillators and filters using permutation equivariant continuous normalising flows, the method outperforms both regression-based approaches and symmetry-naive generative models on both synthetic tasks and a real-world synthesiser. You can read the paper "Audio Synthesizer Inversion in Symmetric Parameter Spaces With Approximately Equivariant Flow Matching" on the conference website. The AIM Centre for Doctoral Training in Artificial Intelligence and Music is funded by RCUK. Based at Queen Mary University of London, AIM students undertake a four year PhD focused on developing cutting-edge research in collaboration with our industry partners.

George Fazekas awarded prestigious Research Fellowship

Sun, 14 Sep 2025 23:00:00 +0100

George Fazekas has been awarded a prestigious Royal Academy of Engineering (RAEng) / Leverhulme Trust Research Fellowship 2025/26. The Fellowship explores smarter AI for music by combining data-driven learning with musical knowledge, making models faster, more efficient, and better at understanding and analysing music. More information can be found at the Royal Academy of Engineering website.