Three Compass CDT students presented their research at the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025) earlier this month, following the acceptance of their papers by the prestigious event.
Held at venues in both San Diego and Mexico City, the world-leading conference on Artificial Intelligence (AI), provided them with a valuable opportunity to highlight their work, and to network with leading figures from both academia and industry.
The students were all lead authors for their papers, one of which received a ‘NeurIPS Spotlight Award’ – with several other researchers from the Institute for Statistical Sciences, including Compass supervisors, cited as co-authors:
- ‘Conditional Distribution Compression via the Kernel Conditional Mean Embedding’
- Dominic Broadbent · Nick Whiteley · Robert Allison · Tom Lovett
- ‘Direct Fisher Score Estimation for Likelihood Maximization’ * NeurIPS spotlight award
-
- Sherman Khoo · Yakun Wang · Song Liu · Mark Beaumont
-
- ‘Scale-invariant attention’
- Ben Anson · Xi Wang · Laurence Aitchison
The opportunity to showcase their research, to develop their network of international contacts, and to learn from expert speakers and participants, was beneficial to all those who attended.
“During the conference I presented my work in a poster session, where I had engaging discussions with a wide range of academic and industry researchers from diverse backgrounds,” said Dominic Broadbent, one of the Compass representatives.
“It was extremely valuable to practise explaining my work to a highly capable audience, some with little direct experience of my research area. I also had the opportunity to discuss my poster with Professor Kenji Fukumizu, a leading figure in my field.
While the strong focus on large language models (LLMs) was not directly aligned with my own research, I learned an enormous amount about the current state of these models, and the broader challenges in reaching artificial general intelligence.
The keynote talks were fascinating, with discussions on how to move beyond LLMs, how to formalise and measure what it truly means for a model to be intelligent, and exciting theoretical work aimed at simplifying and explaining why current models are already so capable.”
Another Compass student, Ben Anson, found several aspects of the conference to be engaging and valuable: “My favourite tutorial on the first day, ‘Autoregressive Models Beyond Language’, pushed me to think carefully about autoregressive model assumptions we inherit from LLMs and how they do, or do not, transfer to non-text modalities.
The oral presentations were also very interesting, particularly ‘Why Diffusion Models Don’t Memorize‘ and ‘On the Closed-Form of Flow Matching’, which offered clear explanations of why diffusion and flow-matching models can generate novel samples rather than reproducing the training set.
It was inspiring to see Professor Rich Sutton (a reinforcement learning icon) share his vision for the future of AI, centred on his OAK architecture (which has an unfortunate dependency on continual deep learning), and to hear Dr Kaiming He’s Test-of-Time talk, where he reflected on how hard training models was twenty years ago and likened research to navigating a ship through fog with no particular destination.
The best part of the conference, however, was the poster sessions. It was great talking with authors, meeting incredibly talented deep-learning researchers, and presenting my own poster – where it turned out that people have plenty of gripes with long-context attention. It was very helpful to discuss some of the little details and issues with people from industry about what works at scale.”
Compass student, Sherman Khoo, who presented his NeurIPS spotlight award-winning paper at the conference, also benefitted from the opportunity to share and discuss his research: “Presenting at the NeurIPS conference provided an excellent opportunity to share our work and obtain feedback from the broader machine learning community.
This was particularly valuable for my research which, while focusing on the statistical problem of maximum likelihood inference in simulation-based inference, had potentially wider applicability, and attracted interest from delegates with interests beyond my immediate research area.
Alongside the exciting, fast-paced discussions in the main conference, I found the workshops to be especially useful for learning about the cutting-edge research across multiple fields. One I particularly enjoyed was the ‘Machine Learning and the Physical Sciences’ workshop, which featured several practitioners applying simulation-based inference across a range of scientific domains, and provided many interesting potential research problems and future directions.”
Three other papers co-authored by members of the Institute for Statistical Sciences, in addition to those involving Compass students, were also accepted by NeurIPS 2025, making it a particularly impactful event for the University of Bristol’s School of Mathematics.
Speaking in advance of the event, Professor Nick Whiteley, Director of the Compass CDT and Head of the Institute, said: “Publishing research in the proceedings of the NeurIPS conference is extremely competitive.
Six acceptances this year illustrates the strength of the Institute for Statistical Science in AI and machine learning, and I’m delighted to see Compass students playing a central role in authorship of these papers.”