Tue 03 Jun 13:00: All-atom Diffusion Transformers: Unified generative modelling of molecules and materials hybrid and recorded
Diffusion models are the standard toolkit for generative modelling of 3D atomic systems. However, for different types of atomic systems – such as molecules and materials – the generative processes are usually highly specific to the target system despite the underlying physics being the same. We introduce the All-atom Diffusion Transformer (ADiT), a unified latent diffusion framework for jointly generating both periodic materials and non-periodic molecular systems using the same model: (1) An autoencoder maps a unified, all-atom representations of molecules and materials to a shared latent embedding space; and (2) A diffusion model is trained to generate new latent embeddings that the autoencoder can decode to sample new molecules or materials. Experiments on QM9 and MP20 datasets demonstrate that jointly trained ADiT generates realistic and valid molecules as well as materials, exceeding state-of-the-art results from molecule and crystal-specific models. ADiT uses standard Transformers for both the autoencoder and diffusion model, resulting in significant speedups during training and inference compared to equivariant diffusion models. Scaling ADiT up to half a billion parameters predictably improves performance, representing a step towards broadly generalizable foundation models for generative chemistry. Open source code: https://github.com/facebookresearch/all-atom-diffusion-transformer
hybrid and recorded
- Speaker: Chaitanya Joshi
- Tuesday 03 June 2025, 13:00-14:00
- Venue: somewhere, Computer Laboratory, William Gates Building.
- Series: Artificial Intelligence Research Group Talks (Computer Laboratory); organiser: Pietro Lio.
Fri 30 May 14:00: Aeneas: Rust Verification by Functional Translation
We present Aeneas, a verification toolchain for Rust programs based on a lightweight functional translation. We leverage Rust’s rich region-based type system to generate pure models for a large class of safe Rust programs which can contain shared and mutable borrows, functions returning borrows, traits and loops. Doing so, we allow the user to reason about the original Rust program through the theorem prover of their choice, and enable lightweight verification by eliminating memory reasoning, allowing them to instead focus on functional properties of their code. As of today, Aeneas has backends for F\*, Coq, HOL4 and most importantly Lean, for which we are investing efforts to develop custom tactics and automation. Aeneas is currently being used to verify optimised, low-level cryptographic code, that we will present as well.
- Speaker: Son Ho (Azure Research team)
- Friday 30 May 2025, 14:00-15:00
- Venue: SS03, Computer Laboratory.
- Series: Logic and Semantics Seminar (Computer Laboratory); organiser: Ioannis Markakis.
Fri 23 May 12:00: Robust Alignment of Large Language Models
The alignment of large language models (LLMs) can often be brittle when faced with the complexities of real-world deployment. In this talk, I share our investigations on two scenarios where special care is required to ensure robust alignment.
The first scenario is multi-objective alignment, where balancing competing objectives is particularly challenging. Our recent work, Robust Multi-Objective Decoding (RMOD), an inference-time alignment algorithm, adaptively adjusts the weights of different objectives during response generation to ensure none are neglected. RMOD provides principled robustness with minimal overhead, consistently outperforming existing methods across several alignment benchmarks.
In the second part of the talk, I will address preference model misspecification in self-play alignment. While self-play is a promising alignment approach, naive implementations are vulnerable to inaccuracies in the preference model. To address this, our Regularized Self-Play Policy Optimization (RSPO) framework offers a versatile and modular method for regularizing the self-play alignment process. RSPO ’s ability to combine various regularizers results in strong performance gains on multiple evaluation sets, such as AlpacaEval-2 and Arena-Hard.
As a bonus, I will briefly introduce our recent investigation into the robustness of Mixture-of-Agent (MoA) systems, a popular multi-agent paradigm. We show that even a single malicious agent introduced into the mixture can nullify the benefits of the entire system.
- Speaker: Dr. Sangwoong Yoon (UCL)
- Friday 23 May 2025, 12:00-13:00
- Venue: ONLINE ONLY. Here is the Zoom link: https://cam-ac-uk.zoom.us/j/4751389294?pwd=Z2ZOSDk0eG1wZldVWG1GVVhrTzFIZz09.
- Series: NLIP Seminar Series; organiser: Suchir Salhan.
Wed 22 Oct 14:30: Title to be confirmed
Abstract not available
- Speaker: Dr Mihai-Cosmin Marinica, CEA Saclay
- Wednesday 22 October 2025, 14:30-15:30
- Venue: Unilever Lecture Theatre, Yusuf Hamied Department of Chemistry.
- Series: Theory - Chemistry Research Interest Group; organiser: Lisa Masters.
Thu 29 May 13:00: Scaling Up Forest Vision with Synthetic Data: Current Progress and Next Steps
Abstract
Machine learning rests on three pillars: algorithms, hardware, and data. In the context of close-range forest monitoring, we’ve already seen major advances in the first two—shifting from classical processing methods to neural networks, and from manual tools like tape measures to LiDAR-based laser scanning. These breakthroughs have enabled the development of faster and more accurate forest monitoring algorithms.
However, data remains a bottleneck. High-quality, annotated forest datasets are scarce and costly to produce, and their size still falls short of the scale required for robust machine learning. Meanwhile, the rise of graphics engines—and the success of synthetic data in domains like self-driving and robotics—makes us wonder: can forests benefit from a similar approach? The key challenge lies in whether synthetic forest environments can capture the representations needed for generalisation to real-world data.
In this talk, I’ll focus on the task of instance segmentation of individual trees—a core bottleneck in many field applications. I’ll present my current progress in generating synthetic forest plots and point cloud data using Unreal Engine, and evaluate their performance against a state-of-the-art model trained on a leading real-world dataset. I’ll also discuss upcoming directions and experimental plans. Time permitting, I’ll give a live demo of my synthetic data pipeline, showing how we can go from video games to ML-ready datasets.
This is a work-in-progress talk, and I look forward to feedback and discussion.
Bio
Yihang She is a second-year PhD student in Computer Science at the University of Cambridge. His research focuses on advancing computer vision in the novel context of forest monitoring, spanning both close-range and satellite-based observations.
- Speaker: Yihang She, University of Cambridge
- Thursday 29 May 2025, 13:00-14:00
- Venue: Room GS15 at the William Gates Building and on Zoom: https://cl-cam-ac-uk.zoom.us/j/4361570789?pwd=Nkl2T3ZLaTZwRm05bzRTOUUxY3Q4QT09&from=addon .
- Series: Energy and Environment Group, Department of CST; organiser: lyr24.
Wed 16 Jul 11:30: Understanding health and wellbeing during menopause: A longitudinal comparison study between autistic and non-autistic people
Autistic people may experience a particularly difficult time during menopause compared to non-autistic people, but there are few direct comparisons as to why this might be. We examined whether menopause representations (thoughts, beliefs and feelings about menopause) are associated with menopausal symptoms, quality of life and mental health in autistic and non-autistic people over a month of time. We found that autistic participants consistently showed negative menopause representations, which predicted more severe symptoms (menopausal, depressive, anxiety) and lower quality of life. These outcomes can contribute to improving awareness of menopause in the autistic population and developing autism-specific support for menopause.
- Speaker: Eunhee Kim, Bournemouth University
- Wednesday 16 July 2025, 11:30-12:30
- Venue: https://us02web.zoom.us/j/87076030035?pwd=XUpJuh8jiR0mae1AhkV79qbg8MtlSM.1.
- Series: ARClub Talks; organiser: Simon Braschi.
Tue 20 May 14:30: The Integral Hilbert Property on some degree-2 del Pezzo surfaces
In this talk I will consider a Diophantine equation known also as “Fermat’s near miss”. Studying the existence and abundance of integral solutions of this kind of equations can be translated in studying the integral points on a degree-2 del Pezzo surface. Rational points on these surfaces have been widely studied, but results on integral points are still missing. I will present a joint work (in progress) with Dan Loughran, where we show that on this surface we have “many” (i.e. a dense set of) integral points.
- Speaker: Jessica Alessandri (Bath)
- Tuesday 20 May 2025, 14:30-15:30
- Venue: MR13.
- Series: Number Theory Seminar; organiser: Jef Laga.
Tue 20 May 14:00: Two Tales of End-to-End Encryption
End-to-end encryption has become the gold standard for securing communications, bringing strong privacy guarantees to billions of users worldwide. My talk will discuss two potential avenues through which the security protections of end-to-end encryption might be undermined, concerning government circumvention and commercial circumvention respectively. First, I examine a recent case of a large-scale law enforcement hack of an encrypted communication network called Encrochat in Europe. Second, I raise security concerns related to the integration of AI features across apps and devices—a growing trend, following remarkable recent advances in generative artificial intelligence.
Zoom link: https://cam-ac-uk.zoom.us/j/82651240291?pwd=ffMwua3b0yUz0C2TFaIxzra2aplAwr.1
- Speaker: Sunoo Park, NYU
- Tuesday 20 May 2025, 14:00-15:00
- Venue: Webinar & LT2, Computer Laboratory, William Gates Building..
- Series: Computer Laboratory Security Seminar; organiser: Anna Talas.
Fri 13 Jun 14:30: Layered Electrides: What They Are, What They Do, and What Can They Do
An electride is a type of ionic material where an electron localized in a void acts in place of an anion. A wide variety of electrides have been synthesized, spanning many stoichiometries and geometries. From a theory perspective, they are of interest for their chemical and material applications, and for their unique electronic structure. A sub-class of electrides are those that are also layered materials, featuring electrons in the interlayer spacing. Modelling these materials with density-functional theory (DFT) requires a sufficient theoretical framework that accounts for dispersion forces and uses a basis set that can capture the electride states. Their large exfoliation energies and small coefficients of thermal expansion make layered electrides unique among the family of layered materials. One application that could leverage the appealing properties of layered electrides is to improve charge transfer across metal/MoS2 heterojunctions. Constructing low-resistance contacts is currently a major challenge for the semiconductor industry, and the insertion of a monolayer electride (Ca2N) to from a metal/Ca2N/MoS2 heterojunction could be the key to an Ohmic contact. Of interest is which metal(s) would be optimal for such a heterojunction.
- Speaker: Adrian Rumson, Dalhousie University
- Friday 13 June 2025, 14:30-15:30
- Venue: Unilever Lecture Theatre, Department of Chemistry.
- Series: Extra Theoretical Chemistry Seminars; organiser: Lisa Masters.
Wed 21 May 11:00: Evaluating and Regulating Foundation Models Teams link available upon request (it is sent out on our mailing list, eng-mlg-rcc [at] lists.cam.ac.uk). Sign up to our mailing list for easier reminders via lists.cam.ac.uk.
The emergence of foundation models and generalist AI systems has transformed the landscape of evaluation, introducing complex challenges that go far beyond the closed-domain settings of the past. This reading group aims to explore cutting-edge approaches for assessing these open-domain systems, with an emphasis on both technical evaluation strategies and evolving regulatory frameworks. We will begin by examining the unique difficulties of evaluating open-domain models, considering possible solutions and highlighting the risks of metric manipulation by resourceful actors. Next, we will discuss the methodologies employed by frontier labs for internal evaluation, as well as the interplay between technical validation and policy-driven oversight. Finally, we will explore evaluation in the context of human-machine collaboration, analyzing the challenges of measuring performance and alignment in systems with humans in the loop.
Teams link available upon request (it is sent out on our mailing list, eng-mlg-rcc [at] lists.cam.ac.uk). Sign up to our mailing list for easier reminders via lists.cam.ac.uk.
- Speaker: Miri Zilka, Neel Alex, Shoaib Ahmed Siddiqui, University of Cambridge
- Wednesday 21 May 2025, 11:00-12:30
- Venue: Cambridge University Engineering Department, CBL Seminar room BE4-38..
- Series: Machine Learning Reading Group @ CUED; organiser: .
Thu 22 May 11:30: Nano in Precision Medicine: Applications in Early Cancer Detection and Drug Delivery
Modern molecular science, best described as the convergence of (bio)chemistry, molecular biology, physics, and the transformative power of artificial intelligence, is paving the way in design of advanced materials for precision medicine. One of the most promising innovations are bio-nano hybrids, engineered systems that seamlessly integrate biomolecules into synthetic nanostructures to bridge the gap between biology and technology.
In this talk, we will delve into the design principles and biomedical applications of organic and hybrid nanomaterials, with a focus on their role in real-time detection of senescent (aged) cells and the role they play in early cancer detection, and targeted drug delivery.
Beyond the current applications, we will also explore how the next generation of bio-nano tools could redefine the future of biomedical nanotechnology, unlocking new possibilities for disease monitoring, intervention, and treatment. We will also examine the key challenges that must be addressed to streamline the translation pipeline from labs to clinic.
- Speaker: Prof Ljiljana Fruk ( Department of Chemical Engineering and Biotechnology)
- Thursday 22 May 2025, 11:30-12:30
- Venue: Open Plan Area, Institute for Energy and Environmental Flows, Madingley Rise CB3 0EZ.
- Series: Institute for Energy and Environmental Flows (IEEF); organiser: Catherine Pearson.
Thu 22 May 13:00: Hydrogen airplanes and why they need Heat Exchangers
Hydrogen aircraft are a promising technology in the long-term for flights beyond what is achievable with batteries. One of the key elements to make an economically competitive H2 aircraft are the heat exchangers, focus of the talk.
- Speaker: Kilian Bartsch, Department of Engineering
- Thursday 22 May 2025, 13:00-14:00
- Venue: Richard King room, Darwin College.
- Series: Darwin College Science Seminars; organiser: Alexander R Epstein.
Thu 30 Oct 14:00: Electronic response to ion projectiles traversing matter from first principles
Abstract not available
- Speaker: Prof. Emilio Artacho (Cambridge & Nanogune)
- Thursday 30 October 2025, 14:00-15:30
- Venue: Seminar Room 3, RDC.
- Series: Theory of Condensed Matter; organiser: Bo Peng.
Thu 04 Jun 17:00: LLMs
Abstract not available
- Speaker: Valeria Ruscio
- Thursday 04 June 2026, 17:00-17:45
- Venue: Lecture Theatre 2, Computer Laboratory, William Gates Building.
- Series: Foundation AI; organiser: Pietro Lio.
Fri 30 May 16:00: TBA
Abstract not available
- Speaker: Hannah Banks (Cambridge U.)
- Friday 30 May 2025, 16:00-17:00
- Venue: Ray Dolby Centre, Seminar Room - North (Floor: 0 A0.019).
- Series: HEP phenomenology joint Cavendish-DAMTP seminar; organiser: Nico Gubernari.
Fri 23 May 08:45: JCTS Presentations
Charli Twyford: Grass Awn Foreign Bodies: Diagnosis and Retrieval in 18 Canine Patients Nicholas Kelly: Pneumoperitoneum in two cats treated with trilostane for hyperadrenocorticism Amy Lord: Antibiotics use in patients seen through a small animal internal medicine referral service
Chaired by Katie McCallum
- Speaker: Charli Twyford, Nicholas Kelly and Amy Lord, Department of Veterinary Medicine
- Friday 23 May 2025, 08:45-10:00
- Venue: LT2.
- Series: Friday Morning Seminars, Dept of Veterinary Medicine; organiser: Fiona Roby.
Fri 30 May 08:45: JCTS Presentations
Harry Peters: Does BOAS surgery reduce the incidence of neurological conditions in later life? Sub title- A brief review of the pitfalls of questionnaire based research. Anna Preston: A case of a subungual keratoacanthoma in a greyhound.
Chaired by Armando Sanchez-Lara
- Speaker: Harry Peters and Anna Preston, Department of Veterinary Medicine
- Friday 30 May 2025, 08:45-10:00
- Venue: LT2.
- Series: Friday Morning Seminars, Dept of Veterinary Medicine; organiser: Fiona Roby.
Wed 28 May 16:00: Typical hyperbolic surfaces have an optimal spectral gap
The first non-zero Laplace eigenvalue of a hyperbolic surface, or its spectral gap, measures how well-connected the surface is: surfaces with a large spectral gap are hard to cut in pieces, have a small diameter and fast mixing times. For large hyperbolic surfaces (of large area or large genus g, equivalently), we know that the spectral gap is asymptotically bounded above by 1/4. The aim of this talk is to present joint work with Nalini Anantharaman, where we prove that most hyperbolic surfaces have a near-optimal spectral gap. That is to say, we prove that, for any ε>0, the Weil-Petersson probability for a hyperbolic surface of genus g to have a spectral gap greater than 1/4-ε goes to one as g goes to infinity. This statement is analogous to Alon’s 1986 conjecture for regular graphs, proven by Friedman in 2003. I will present our approach, which shares many similarities with Friedman’s work, and introduce new tools and ideas that we have developed in order to tackle this problem.
- Speaker: Laura Monk (Bristol)
- Wednesday 28 May 2025, 16:00-17:00
- Venue: MR13.
- Series: Differential Geometry and Topology Seminar; organiser: Oscar Randal-Williams.
Thu 22 May 13:00: Compute Cluster and Storage
Abstract
Designing and building a distributed processing cluster of GPUs and CPUs with petabytes of storage, geared towards processing environmental data.
Bio
Mark is a Senior Research Software Engineer and Tarides Fellow in the Energy and Environment Group. His areas of interest are the OCaml ecosystem, large-scale distributed computing clusters.
- Speaker: Mark Elvers, University of Cambridge
- Thursday 22 May 2025, 13:00-14:00
- Venue: Room GS15 at the William Gates Building and on Zoom: https://cl-cam-ac-uk.zoom.us/j/4361570789?pwd=Nkl2T3ZLaTZwRm05bzRTOUUxY3Q4QT09&from=addon .
- Series: Energy and Environment Group, Department of CST; organiser: lyr24.
Fri 31 Oct 14:00: Title to be confirmed
Abstract not available
- Speaker: Dr Shelly Zhang, University of Illinois at Urbana-Champaign
- Friday 31 October 2025, 14:00-15:00
- Venue: Oatley 1 Meeting Room, Department of Engineering.
- Series: Engineering - Mechanics and Materials Seminar Series; organiser: div-c.