MARS researchers have had seven papers accepted to the main conference of NeurIPS 2026, one of the leading international venues for machine learning research. A further six papers will appear at the conference’s specialist workshops.

NeurIPS, which attracts more than 20,000 attendees, takes place in Sydney from 6–12 December, with satellite events in Paris and Atlanta from 9–13 December.

“This is an exceptional achievement for the MARS research group, which was only established two years ago,” said Professor Chris Nemeth, Director of MARS Research.

“The acceptance of so many papers demonstrates the excellence of MARS AI research, the talent of our researchers, many of whom are at an early stage in their careers, and our growing standing in the international machine learning community.”

The seven main conference papers cover several of MARS’s core themes. These include Gaussian processes, Bayesian optimisation, diffusion-based generative modelling, active learning and uncertainty quantification.

Gaussian processes

Gaussian processes (GPs) estimate an unknown, continuously varying quantity from limited measurements, giving both a best estimate and a measure of uncertainty. However, they produce exact results only when measurements relate to the quantity in a simple, proportional way and their errors follow a bell-shaped (normal) distribution.

In Conditioning Gaussian Processes on Almost Anything, Henry Moss, Lachlan Astfalck, Thomas Cowperthwaite, Colin Doumont, Sam Willis, Philipp Hennig, Chris Nemeth and Andrew Zammit-Mangion remove these restrictions. They show that GPs are mathematically equivalent to a class of diffusion models, the technology underpinning many AI image generators. Their method reproduces standard GP results exactly and can use almost any type of information, as long as there is a way to measure how consistent a possible prediction is with that information. This includes complex physical laws and statements in natural language, which can be interpreted with the help of large language models.

In Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights, Jixiang Qing, Henry Moss and Matthias Sachs tackle a challenge from chemistry: mapping how a system’s energy varies with the arrangement of its atoms, when each calculation is computationally expensive to run. Their method selects the most useful calculation to run next, focusing effort on the arrangements that matter most while still exploring areas where the model is uncertain. The authors prove that the method’s accuracy reliably increases as more data is collected, and show improvements over existing methods on molecular modelling and drug discovery tasks.

Bayesian optimisation

Bayesian optimisation searches for the best option when each test is expensive, using the results so far to choose the most promising next test.

In Search at the Cost of Sampling: Nearly-Instant Latent Space Bayesian Optimization, Donney Fan, Colin Doumont, Aleksandra Kalisz, Paul Duckworth, Jacob Gardner, Henry Moss and Geoff Pleiss focus on AI design pipelines where new designs can be checked quickly, so the bottleneck is choosing which design to try next. By pairing a simple model with the way these AI systems store designs, clustered on the surface of a multidimensional sphere, they make the search at least 100 times faster than the best existing methods, while finding designs that are as good or better.

In BO-Arena: An Evolving Benchmark for High-Dimensional Bayesian Optimisation, Thomas Christie, Colin Doumont, Henry Moss and Philipp Hennig introduce a shared platform for testing and comparing Bayesian optimisation methods on problems involving a large number of variables. It is designed to grow over time as new problems and methods are added.

Diffusion models

Diffusion models create content by gradually refining random noise into a structured result.

In Tempered Guided Diffusion, Andreas Makris, Paul Fearnhead and Chris Nemeth use diffusion models to reconstruct images from incomplete or indirect information. Instead of fully generating several candidates and keeping the best, their method checks candidates partway through the process and fully generates only the most promising. The result is generally more accurate reconstructions for the same computing time, and a better balance between speed and quality.

In Hypergraph Generation via Structured Stochastic Diffusion, Chris Nemeth introduces HEDGE, which generates hypergraphs: networks of relationships among groups rather than pairs. Many approaches break groups into pairs, losing information about group sizes and overlaps. HEDGE works with groups directly, is mathematically proven to keep errors under control, and reproduces the way groups overlap more faithfully than existing methods.

Uncertainty quantification

Uncertainty quantification measures how reliable a model’s predictions are. One widely used approach, conformal prediction, produces ranges guaranteed to contain the true answer at least a specified proportion of the time, provided its reference data comes from the same source as new data.

In When Does Trimming Help Conformal Prediction? A Retained-Law Diagnostic under Calibration Contamination, Congye Wang examines trimming: removing suspicious-looking points from contaminated reference data. The paper shows that whether trimming helps depends on the data that remains, not simply on how much was contaminated. It offers a way to judge when removing data is worthwhile and when it may weaken the reliability of predictions.

The main conference papers will appear in the NeurIPS 2026 proceedings, with most preprints available on arXiv. A list of workshop papers is available below.

Workshop papers

• C4-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation: Lipinski, D., Qing, J. and Moss, H.
• Topology-Informed Active Learning for Sequential Materials Screening: Cao, Y. and Qing, J.
• Conditioning Stochastic Processes using Guided Latent Dynamics: Sharrock, L., Astfalck, L. and Moss, H.
• Scalable Gaussian Process Flows: Cowperthwaite, T., Sharrock, L., Astfalck, L. and Moss, H.
• How Reliable are Intermediate Rewards for SMC-based Guidance in Masked Discrete Diffusion Models? A Closed-Form Analysis: Christie, T., Ek, C., Moss, H. and Hennig, P.
• Convergence of Potential-Preserving Annealed Sinkhorn under Non-Summable Temperature Schedules: Lipinski, D. and Buze, M.