Last week, we welcomed doctoral students from across the UK for a Turing PhD Connections event – a day dedicated to collaboration and conversation whilst exploring cutting-edge research in data science and artificial intelligence.

• Develop a more cohesive community of doctoral students across organisations,
• Enhance student skills and research practices beyond their individual PhD programmes, and
• Provide opportunities to engage with real-world challenges and applications of data science.

• Trust in human-robot interaction,
• Automatic assessment of fine-grained levels of pain and related states in chronic pain using multimodal signals, and
• The effects of different fact-checking interventions on engagement with political misinformation among Republicans.

Poster Session

Last month, students, researchers, and industry leaders gathered for an inspiring three-day CDT Showcase to explore cutting-edge developments in artificial intelligence, machine learning, and their applications across society. The event combined academic talks, poster sessions, interactive activities, and social events, creating a vibrant environment for learning and collaboration.

The Summer School opened with a lively icebreaker and networking session, setting the tone for three days of exchange and exploration.

Prof Marc Deisenroth

A thought-provoking discussion chaired by Prof Marc Deisenroth examined AI for Environment and Sustainability. Following a presentation by Marc, delegates were divided into groups and tasked with exploring and reporting back on topics, including ‘Where do you see potential for AI transformation in environmental modelling’ and ‘Where do you see potential for AI transformation in sustainable energy’. This was a great opening session, encouraging participants to reflect on the role of technology in tackling global challenges!

Later, keynote speaker Dave Palfrey, Chief Scientist at MindMage gave a fascinating talk on code modularity and division of labour,  drawing parallels between human collaboration and the way we design AI systems. The final talk was CDT researcher David Chanin’s work on improving sparse autoencoders.

Dave Palfrey

The day concluded with a poster exhibition showcasing the breadth of student research, before delegates headed to an escape room social event in central London – a perfect chance to test teamwork skills outside of the showcase!

UCL East’s Intelligent Robotics Lab

The final day featured an exclusive tour of UCL East’s Intelligent Robotics Lab, where our guests learned how AI and robotics are being applied to fields such as autonomous vehicles, sustainable agriculture, and industrial inspection.

The afternoon’s talks tackled major questions around the future of AI research. Highlights included Reuben Adams on AI safety and the automation of AI research, Laura Ruis on how large language models learn reasoning skills, and

Prof. Tim Rocktäschel

Prof. Tim Rocktäschel with his ICLR keynote on open-endedness and the automation of innovation. These sessions pushed participants to think critically about the pace of AI development and its long-term implications.

Looking Ahead

The 2025 CDT Showcase was a testament to the energy, creativity, and vision of the AI research community. From student-led breakthroughs to expert insights on sustainability, safety, and innovation, the event highlighted both the opportunities and responsibilities that come with shaping the future of AI.

We look forward to continuing these conversations and seeing where the next year of research will take us.

Best Poster and Programmed Art Prizewinners

Over the past year, my research has focused on the following questions: how do large language models (LLMs) learn to do reasoning from large-scale textual pretraining? Anticipating the eventual surpassing of human experts by AI, how can we still make sure that what they tell us makes sense? And, given that the field is using LLMs as evaluators in most papers nowadays, how can we make sure their evaluations are reliable? In this post, I will go over me and my collaborators past year of findings.

How Do Models Learn to Reason from Pretraining Data?
My main research output came with a paper titled “Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models,” which was accepted at ICLR 2025. Since the advent of LLMs and their saturation of benchmarks, a question has been bugging me: are models truly reasoning when they solve problems, or have they merely seen very similar questions as the ones we benchmark them with in their vast, internet-scale training data? Tackling this question was challenging for a few reasons. Firstly, the pretraining data is so vast that searching through it is usually intractable (trillions of tokens!). Secondly, doing interpretability to understand what pretraining data influences model outputs is very expensive, as models have billions of parameters.

To answer this question, I investigated how the training data models rely on when reasoning differs from the training data they rely on when doing factual question answering. The hypothesis here is as follows: if models are memorising answers to reasoning questions, the patterns of influential pretraining data for reasoning may look quite similar to the patterns for factual question answering. What I discovered was fascinating: while models rely on completely different data sources when answering different factual questions (like “What’s the capital of France?” and “What is the highest mountain in the world?”), they consistently draw from the same data when solving different reasoning problems within the same category. This suggests they can learn from the same pretraining data and apply what they’ve learned to many different reasoning questions, indicating the presence of procedural knowledge. For the factual question answering, I found pretraining data with the answers to the questions to be highly influential, whereas for the reasoning questions that was not the case. My findings suggest that AI models aren’t just retrieving memorised answers—they’re learning procedural knowledge, like how to apply formulas or follow problem-solving steps.

Think of it like the difference between memorising that 2+2=4 versus learning the procedure for addition that can be applied to different numbers. My research shows that AI models are doing more of the latter than previously thought, which has important implications for how we understand their capabilities and limitations.
More about this work in my blogpost.

Making AI Debates More Truthful
At ICML 2024 in Vienna, collaborators and me presented work on an entirely different but equally important problem: as models grow increasingly sophisticated, they will surpass human expertise, and the role of humans and AI will flip. Humans will need to start evaluating AI, as opposed to vice-versa. But how can we evaluate AI that is smarter than us? Our paper “Debating with More Persuasive LLMs Leads to More Truthful Answers” takes a first step in this direction using a creative solution inspired by human debate, and even won a best paper award!

The concept is simple: instead of relying on humans to directly evaluate AI responses, we have two AI systems debate different answers to a question, then let a third AI (or human) judge which argument is more convincing. Remarkably, this approach helped both AI judges and humans identify correct answers 76% and 88% of the time respectively, compared to much lower accuracy with simpler methods.

Even more intriguingly, when we used inference-time methods to make the debating AI systems more persuasive, it actually made them better at revealing the truth rather than misleading judges. This suggests that we could use AI debate to evaluate their truthfulness as they become more capable, and eventually surpass human expertise.

Improving How We Evaluate AI Systems
The most rewarding experience this past year has been advising students, one of which was Yi Xu, a UCL master’s student. Together with Robert Kirk, we tackled a crucial question of how we reliably evaluate and compare different AI systems. Yi developed new methods for ranking AI chatbots that will be presented as a spotlight paper at ICML this year.

The problem we tackled is surprisingly common: current evaluation methods often produce inconsistent rankings, where System A beats System B, System B beats System C, but System C somehow beats System A. The solution draws inspiration from sports tournaments, using round-robin style comparisons combined with statistical modeling to produce more reliable rankings while reducing computational costs.

Advancing AI Safety and Interpretability
Another student I advised through a collaboration with the University of Michigan was Itamar Pres, who works actively on advancing the broader field of AI safety and interpretability. At NeurIPS 2024 in Vancouver, Itamar presented his spotlight paper titled “Towards Reliable Evaluation of Behaviour Steering Interventions in LLMs,” at the MINT workshop. This research focuses on how we can reliably modify AI behaviour and measure the effectiveness of our interventions.

Personally, being at NeurIPS in Vancouver also provided me with the amazing opportunity to share insights with a broader audience through an invited interview on Machine Learning Street Talk, a popular science podcast. The discussion, titled “How do AI models actually think,” allowed me to communicate the implications of my research to both technical and general audiences. My appearance on Machine Learning Street Talk received more than 30K views! How Do AI Models Actually Think?

Looking Forward
This year’s research has reinforced my belief that understanding AI systems requires looking beyond their outputs to examine how they actually process information and make decisions. Whether it’s investigating what training data influences their reasoning, developing better evaluation methods, or creating systems that can reliably identify truth through debate, the work emphasizes transparency and reliability in AI development.

After I wrap up my PhD at UCL this year, I will be joining Professor Jacob Andreas at MIT for a postdoc, and I’m excited to build on these insights and tackle the next generation of challenges in AI alignment and interpretability. I look back on a happy PhD and am grateful I got the chance to pursue this research at UCL.

I recently had the pleasure of attending the 2025 Nengo Neuromorphic AI Summer School, held at the University of Waterloo in Ontario, Canada. The program was expertly organized by Prof. Chris Eliasmith, Dr. Terry Stewart, and Dr. Michael Furlong.

Neuromorphic AI Fundamentals
The first 3 days of the summer school were dedicated to lectures on the fundamental concepts of neuromorphic AI, delivered by world-leading experts from the Centre for Theoretical Neuroscience. This foundational knowledge was crucial for tackling the hands-on research projects that followed.

A central topic was the Spiking Neural Network (SNN), which differs significantly from traditional Artificial Neural Networks (ANNs). While ANNs process continuous numerical signals synchronized by a global clock, SNNs communicate using discrete, time-encoded events called “spikes”—a closer approximation to biological neural activity.

SNNS offer significant advantages in terms of energy efficiency, particularly when deployed on dedicated neuromorphic hardware such as Intel Loihi, SpiNNaker, or Braindrop. As a researcher at the intersection of AI and robotics, I find this efficiency especially promising for applications on resource-constrained systems like mobile robots. This could pave the way for low-power, on-device, open-ended learning systems.

The Nengo Framework is a powerful, flexible tool for designing and deploying large-scale spiking neural models. It serves as a bridge between high-level computational goals and low-level neural dynamics. Built upon the Neural Engineering Framework (NEF) developed by Prof. Eliasmith and colleagues, Nengo allows users to define desired functions, and optimise for the synaptic weights needed for a population of spiking —typically leaky integrate-and-fire (LIF) neurons with random tuning curves—to approximate those functions. Nengo also supports multiple backends, enabling users to run models on standard CPUs and GPUs or deploy them directly to neuromorphic chips for real-world, low-power applications. Lectures throughout the week also introduced more advanced topics such as the Legendre Memory Unit (LMU), Semantic Pointer Architecture (SPA), adaptive spiking neural controllers, and neuromorphic SLAM, among others.

Summer school research project
The remainder of the summer school was dedicated to short research projects. For my project, I proposed and developed a spiking neural controller for quadruped robot locomotion, implemented using Nengo and MuJoCo. My project focused on integrating components such as SNNs, the LMU, spiking Central Pattern Generators (sCPGs), and imitation learning techniques to achieve biologically inspired locomotion controller.

The program concluded with a showcase event, where participants presented their projects in a technical  session and a public demonstration.

To my great delight, my project “SNN for Legged Locomotion Control” was awarded “The Recurrent Award for Best Use of Recurrent Neural Networks” during the closing banquet and award ceremony.

Reflections and Future Work
The Nengo Summer School was a truly transformative experience. The combination of expert-led theoretical sessions, hands-on tutorials, and intensive mini-projects provided a deep and practical understanding of neuromorphic AI. The experience also helped me establish valuable connections for future collaborations, particularly for continuing my research on neuromorphic AI and robotics at UCL.
I highly recommend the Nengo Summer School to anyone with an interest in neuromorphic computing or biologically inspired AI. It’s a rare and enriching opportunity to engage deeply with a cutting-edge field alongside leading researchers and fellow enthusiasts.

On 15 May 2025, a stream of researchers and students wound their way into the Denys Holland Lecture Theatre at UCL, drawn by a shared curiosity: how do we learn reliably when our models are imperfect? This two-day gathering, the inaugural Workshop on Advances in Post-Bayesian Methods—organised by Dr. Jeremias Knoblauch, Yann McLatchie, and  Matías Altamirano (UCL) explored advances beyond the confines of classical Bayesian inference.

Traditional Bayesian methods hinge on having the “right” likelihood and a fully specified prior, then performing a precise update when data arrive. But what happens when those assumptions crumble? In fields from cosmology to epidemiology, models are often approximate, priors are chosen more out of convenience than conviction, and exact computation is out of reach. The answer, as highlighted by the organisers, lies in a broader view of Bayes—one that replaces the rigid likelihood with flexible loss functions or divergences, yielding posteriors that behave more like tools in an optimizer’s kit than tenets of statistical doctrine. Over two days in May, five themes emerged:

1. Reweighting for Robustness
   A number of talks explored how reweighting the data can help account for model misspecification. Ruchira Ray presented statistical guarantees for data-driven tempering, while Prof. Martyn Plummer discussed Bayesian estimating equations leading to inferences which are made invariant to the learning rate.
2. Real-World Impact and Scientific Applications
   Speakers like Devina Mohan and Kai Lehman grounded the discussion in high-impact domains. From galaxy classification to cosmological modeling, these talks showed how post-Bayesian methods are being applied where models are inevitably approximate and uncertainty is essential.
3. Variational Inference at the Forefront
   Variational methods continued to evolve beyond classical forms. Dr. Kamélia Daudel, Dr. Diana Cai, and Dr. Badr-Eddine Cherief-Abdellatif presented advances in black-box inference and importance weighting, illustrating how variational approaches are expanding to handle more structure, complexity, and real-world constraints.
4. PAC-Bayesian Perspectives on Generalization
   PAC-Bayes theory offered a unifying language for understanding how well models generalize. Talks by Prof. Benjamin Guedj and Ioar Casado-Telletxea examined comparator bounds and marginal likelihoods through a PAC-Bayesian lens—providing rigorous guarantees even in adversarial or data-limited regimes.
5. Predictive Bayesian Thinking
   Prof. Sonia Petrone and others emphasized a shift toward prediction-focused Bayesian inference, where the goal is not merely to estimate parameters, but to make useful, calibrated forecasts. This view reframes classical Bayesianism into a pragmatic framework centered on learning what matters.
6. Gradient Flows and Computational Tools
   Finally, computation was treated not as an afterthought but as a core conceptual tool. Dr. Sam Power and Dr. Zheyang Shen discussed using gradient flows and kernel methods to structure inference, showcasing how modern optimization techniques are reshaping the Bayesian workflow itself.

Last week, our PhD students and academic staff traded classrooms and computer screens for countryside views and conversations as we embarked on a three-day retreat designed to recharge and reconnect. The retreat offered a welcome break from daily routines and delivered a memorable mix of insightful talks, friendly competition, reflective workshops, and some great food.

Here’s a look back at the highlights.

The retreat kicked off with a warm welcome and a delicious lunch that set the tone for what would be an engaging and collaborative few days. With everyone gathered in one place, our first structured activity was a team-building session designed to help us connect beyond the confines of our usual roles. Through creative problem-solving and small-group activities, we challenged ourselves to think outside of the box, communicate more effectively, and collaborate in ways that pushed us beyond our comfort zones!

Later in the afternoon, we transitioned into a series of student-led talks. PhD students presented snapshots of their research, prompting engaging discussions and questions.  

The day wrapped up with a fabulous dinner and an informal but highly entertaining evening comprising of impromptu chat and friendly rivalry over board games, breaking the ice even further and setting a welcoming tone for the retreat. 

We began Day 2 with another round of student talks. Building on the previous day, the presentations were dynamic, insightful, and spurred plenty of discussion. The talks showcased the incredible breadth of work happening within our centre and it was  inspiring to see the students speak about their work.

Lunch gave us another chance to unwind and continue conversations sparked during the talks. Then, it was time for the second team-building activity, a treasure hunt around the local village that quickly became a highlight of the retreat.

Armed with clue sheets, maps, and a healthy sense of competition, teams set off on foot to explore the village, solving riddles and deciphering hints that led them to landmarks, a cemetery (!) and hidden details. The experience was packed with laughter, a bit of friendly competition, and plenty of teamwork. Regardless of who “won,” the real reward was the sense of connection and discovery that came with each solved clue.

Back at the retreat venue, dinner that evening felt like a celebration of everything we’d already achieved together. Conversations stretched into the night, blending personal stories with professional ideas. Whether discussing research struggles or favourite hobbies, everyone was enjoying the rare chance to connect informally and meaningfully.

Our final day featured something a little different: a Lego Serious Play workshop – a guided session using Lego bricks to express ideas, metaphors, and visions for our academic and personal journeys. This hands-on session invited participants to use Lego bricks to build metaphors and models that reflected their academic experiences, personal growth, and future aspirations. What might sound whimsical on the surface turned out to be deeply reflective and was a powerful way to close the retreat: creative, thoughtful, and uniquely personal.

Following a final lunch together, we said our goodbyes, refreshed, inspired, and with a renewed appreciation for the people who make up our academic community.

The retreat was more than just a break from routine. It was a space for conversation, connection, and creativity. We came away with new ideas, stronger relationships, and a lego pack! To everyone who made it happen-thank you. Here’s to more moments like these in the future.

I recently had the pleasure of presenting our work “Private Selection with Heterogeneous Sensitivities” at the 3rd IEEE Conference on Secure and Trustworthy Machine Learning (SaTML2025) in Copenhagen. This work was presented in the form of a poster as well as a talk at the main conference.

• We’re working on further improving the adaptive mechanism — several exciting open questions remain.
• We’re also exploring the connection to online learning, especially in settings with distribution shift. We only hinted at this in the paper, but early results are promising and suggest real benefits from private selection in sequential decision-making.

I attended the International Conference on 3D Vision (3DV) in March. 3DV is a conference that brings together researchers in 3D computer vision and graphics. At the conference, I presented our paper, Spann3R: 3D Reconstruction with Spatial Memory, which was accepted by 3DV as an award candidate.

Our work addresses the problem of dense 3D reconstruction from images only, without any prior knowledge of camera parameters or depth information. Due to the inherent ambiguity of interpreting 3D structures from 2D images, traditional dense reconstruction pipelines typically decompose the task into a series of minimal problems, which are then chained together. This approach usually requires substantial manual tweaks as well as engineering efforts. In contrast, our approach directly maps a set of 2D images to 3D using a neural network, enabling a fully feed-forward dense reconstruction.

The key idea of our method is to manage a spatial memory that stores previous states and learns to query this spatial memory to reconstruct the geometry of the subsequent frames. Our results demonstrate a proof-of-concept for feed-forward 3D reconstruction, and open up exciting directions for fully learning-based approaches to Structure-from-Motion (SfM) and Simultaneous Localization and Mapping (SLAM). Interestingly, as a by-product, our method can perform dense reconstruction without explicitly estimating camera poses—a concept that some later works have referred to as pose-free reconstruction.

Beyond my own presentation, I attended several insightful talks, including those by Jon Barron, Noah Snavely, and Fei-Fei Li. In Jon’s talk, he mentioned one particular interesting topic on “Why care about 3D?”. He discussed various perspectives on the role of 3D in computer vision and robotics.

Jon challenged the traditional assumption that a robot must reconstruct its environment in 3D to navigate it effectively. With the progress of deep learning, robots can map observations directly to control signals without explicit 3D modeling—similar to trends in autonomous driving. Likewise, the idea that realistic image or video generation requires full 3D reconstruction is challenged with the rapid progress of video diffusion models.

Jon offered several interesting arguments for why 3D still matters. One reason is efficiency—once a 3D scene is reconstructed, it can be rendered from many viewpoints at low cost. More fundamentally, humans live in a 3D world—so if we expect AI systems to perceive and interact like us, they must reason in 3D too.

Yet, the community has not reached a consensus on how to let AI models learn 3D efficiently. While vast amounts of 2D image data are readily available thanks to smartphones and the Internet, acquiring large-scale, high-quality 3D data remains costly. Hopefully, the progress in 3D foundation models can unlock new ways to scale up the learning of 3D representations from web-scale casual videos without ground-truth 3D, and bridge the gap between 2D and 3D through unified multi-modal representations.

Further information on Hengyi’s project can be found here https://hengyiwang.github.io/projects/spanner

Last month, CDT students travelled to Vancouver to present their work at NeurIPS, one of the largest AI conferences. The schedule was packed, with 6 conference papers and 3 workshop papers presented. The CDT is proud to sponsor and support our PhD students as they present their research at the conference, showcasing their hard work and academic excellence on an international stage!
Day 1, Tuesday: No papers were presented, but the expo, tutorials, and careers fair kept everyone occupied.

Day 2, Wednesday:

William Bankes William and his supervisor presented their work on robust data down sampling. Naive approaches to training on a subset of data can cause problems when classes in the dataset are imbalanced, with rare classes becoming even rarer. A direct result of this is the model’s ability to predict on this subset of the data decreases. To address this they proposed an algorithm called REDUCR, which downsamples data in a manner that preserves the performance of minority classes within the dataset. They show REDUCR works across a range of text and image problems achieving state of the art results. The work is available here.

Jake Cunningham also presented his work on 
Reparameterized Multi-Resolution Convolutions for Long Sequence Modelling

Day 3, Thursday
Yuchen Zhu 

Yuchen’s joint paper with Jialin Yu and Ricardo Silva from UCL Statistical Sciences, Structured Learning of CompositionalSequential Models was presented as part of the main proceedings of NeurIPS 2024. They proposed an explicit model for expressing how the effect of sequential interventions can be isolated into modules, clarifying previously unclear data conditions that allow for the identification of their combined effect at different units and time steps. The paper is here.  Additionally, together with collaborators from MPI Tuebingen, Yuchen presented a paper Unsupervised Causal Abstraction at the Causal Representation Learning workshop. Due to the lack of interpretability of current large blackbox models, they propose a methodology for causally abstracting a large model to a smaller and more interpretable model. In particular, unlike existing methods, their method does not require supervision signals from the smaller model. The paper can be found  here.

David Chanin

David presented a paper along with co-author Daniel Tan, another UCL PhD Student. They find that Contrastive Activation Addition (CAA) steering has mixed results in terms of robustness and reliability. Steering vectors tend to generalise out of distribution when they work in distribution. However, steerability is highly variable across different inputs: depending on the concept, spurious biases can substantially contribute to how effective steering is for each input, presenting a challenge for the widespread use of steering vectors. While CAA is effective on some tasks, other behaviours turn out to be unsteerable. As a result, it is difficult to ensure they will be effective on a given task of interest, limiting their reliability as a general alignment intervention. The paper is available here.

Day 4, Friday
Reuben Adams

Reuben presented his paper (co-authored with his supervisors) on extending a classic theorem in the PAC-Bayes literature to account for arbitrary outcomes, rather than simply correct or incorrect classifications. Their work provides theoretical insight into the generalisation behaviour of neural networks and the different kinds of errors they can make. Their framework can cover not just Type I and Type II errors, but any kind of error that may occur in multi class classification. You can find the paper here.

Daniel Augusto
Daniel presented a co-authored paper in collaboration with
Getúlio Vargas Foundation for the main conference track. Their work proposes a new solution to streaming variational Bayesian inference using  GFlowNets as a foundation for their methodology. This was the first work that allows high quality variational inference for discrete parameters without requiring the storage of the whole dataset. They believe this work will be useful for applications in genetics, through the inference of phylogenetic trees, for preference learning, and other big-data contexts. Their paper can be read here.

Day 5, Saturday 

Oscar Key Oscar, along with co-authors from Graphcore presented a poster at the workshop on Efficient Natural Language and Speech Processing. Their work considers the top-k operation, which finds the largest k items in a list, and investigates how it can be computed as quickly as possible on the parallel hardware commonly used to run AI applications. Top-k is found in many AI algorithms, so it’s useful to make it fast, e.g. a large language model might use it to select the most important parts of a long prompt. Their full paper is available here.

Varsha Ramineni 

Varsha presented her research at the Workshop on Algorithmic Fairness through the Lens of Metrics and Evaluation (AFME 2024). This work addresses the challenge of evaluating classifier fairness when complete datasets, including protected attributes, are inaccessible. They propose a novel approach that leverages separate overlapping datasets, such as internal datasets lacking demographic information  and external sources like census data, to construct synthetic test data with all necessary variables. Experiments demonstrate that our approach produces synthetic data with high fidelity and offers reliable fairness evaluation where real data is limited.  Varsha says that she had an incredible experience attending her first NeurIPS and presenting her work and engaging in meaningful discussions throughout the conference was a deeply rewarding experience, providing invaluable feedback and ideas as the work extends. Do reach out to her if you’d like to learn more!