AI for Global Goals | OxML

OxML 2027: Foundation Models — June, online. Biomedical AI — 10–12 Aug. Frontier AI — 13–15 Aug. Oxford + online. Registration open.

[AI for Global Goals | OxML @GlobalGoalsAI]

OxML 2027 is now open for registration.

Foundation Models — June 2027, online only
Biomedical AI — 10–12 August 2027, Oxford + online
Frontier AI — 13–15 August 2027, Oxford + online

Register: https://t.co/jVliumI7uw

[AI for Global Goals | OxML @GlobalGoalsAI]

🎥 OxML 2026 Speaker Interview: Tim Rocktäschel

How can AI move beyond learning from existing data to making entirely new discoveries?

Tim Rocktäschel, Professor of AI at UCL and Co-founder of Recursive, shares insights on open-ended learning, world models, and how AI could automate innovation by enabling agents to make and build upon new discoveries.

Interested in exploring the frontiers of AI and the latest advancements in the field? Join us at OxML 2027 to learn from leading researchers and industry professionals.

🎓 Explore the 2027 programmes: https://t.co/FYj3ckoZEU

#OxML #FrontierAI #Worldmodels #Openendedness

@ucl @Recursive_SI

[AI for Global Goals | OxML @GlobalGoalsAI]

A benchmark score describes performance under the conditions of that benchmark.

A benchmark measures behaviour under defined conditions. Changed instructions, unexpected tool responses or longer tasks can expose behaviour that a single score leaves untested.

Frontier AI at OxML 2027 covers evaluation pipelines, observability and reliability alongside reasoning and agency. For agent evaluation, the conditions tested matter as much as the reported result.

Frontier AI — 13–15 August 2027 · Oxford + online

Register for Frontier AI: https://t.co/QGUVTF1oPR

#OxML2027

[AI for Global Goals | OxML @GlobalGoalsAI]

Learning the field is different from learning it alongside the people working on it.

Participants in the 2026 Representation Learning & Generative AI programme spent a week at the Mathematical Institute in Oxford, working through frontier models, agents and generative AI. Hear one of them describe the experience.

Frontier AI at OxML 2027 covers learning, reasoning, agency and self-improvement, alongside the engineering behind reliable AI systems.

Join us to learn from peers, compare approaches and discover research that connects to your own.

Frontier AI: 13–15 August 2027 · Oxford + online
Register for Frontier AI: https://t.co/QGUVTF1oPR

#OxML2027 #FrontierAI #MachineLearning

[AI for Global Goals | OxML @GlobalGoalsAI]

For an AI agent, the route to an answer deserves scrutiny alongside the answer itself.

An agent can take questionable steps while planning or using tools and still produce a plausible response. Assessing capability means inspecting actions, evidence and error handling throughout the task.

Frontier AI at OxML 2027 covers learning, reasoning, agency and self-improvement, alongside the engineering behind reliable AI systems.

Frontier AI — 13–15 August 2027 · Oxford + online

Register for Frontier AI: https://t.co/17WTwxlED4

#OxML2027

[AI for Global Goals | OxML @GlobalGoalsAI]

Sequence, structure and function are related modelling problems, but none is a substitute for evidence about the others.

Moving from genomic sequence to protein behaviour involves relationships across biological scales. A model's predictions need biological evidence; a useful representation is not automatically a mechanistic explanation.

Biomedical AI at OxML 2027 covers genomics and multi-omics, proteins and proteomics, structure and function, and translation across scales.

Biomedical AI — 10–12 August 2027 · Oxford + online

Register for Biomedical AI: https://t.co/17WTwxmcsC

#OxML2027

[AI for Global Goals | OxML @GlobalGoalsAI]

Foundation Models | OxML 2027
How are foundation models built, trained and evaluated? And how do their underlying principles extend beyond language?

OxML 2027 explores representation learning, transformers, autoregressive and masked modelling, diffusion models, tokenisation, scaling laws and pre-training.
These foundations are then connected to applications across different domains.

📅 June 2027 | Online
Registration is now open: https://t.co/4rbfIn2nL1

#OxML2027 #FoundationModels

[Mona Alinejad @monalinejad]

Every year, we evolve the OxML programme to reflect the latest advances in AI and machine learning.
For OxML 2027, we’re making some of the biggest changes since the school began.

Our commitment remains the same: advanced AI training taught by leading scientists and researchers.
The new structure reflects how we see some of the major shifts in the field:
• Foundation models are expanding far beyond language and vision.
• Frontier AI increasingly involves systems that reason, use tools, act and improve.

These shifts have shaped three complementary programmes:
1. Foundation Models — Fundamentals & Applications
2. Biomedical AI — Health, Biology & Discovery
3. Frontier AI — Learning, Reasoning, Agency & Self-Improvement

Each programme stands on its own. Together, they offer a connected journey through modern AI, combining theory with applications and hands-on cases.
More on the curriculum and speakers soon:
https://t.co/OlvdX3bXXi
#OxML2027

[AI for Global Goals | OxML @GlobalGoalsAI]

Expanding access to OxML

Excellent researchers and students are distributed globally. Access to advanced training and research communities is not.

In 2026, we continued to expand the financial support available to OxML participants through scholarships, reduced fees for participants from low- and middle-income countries, and dedicated crisis-access support.

This included:

🎓 223 scholarships awarded
🌍 40% support for eligible participants from low- and middle-income countries
🕊️ 100% online access through dedicated crisis scholarships for participants affected by conflict

The objective is simple: financial circumstances or geography should not unnecessarily exclude talented people from opportunities to learn and participate in the research community.

There is considerably more to do, and expanding access will remain an important part of how we develop OxML in 2027.

#OxML2026 #OxML #Scholarships #AccessToEducation #MachineLearning

[AI for Global Goals | OxML @GlobalGoalsAI]

Inside OxML 2026: Oxford's Mathematical Institute

OxML takes place in the Andrew Wiles Building, home of the University of Oxford's Mathematical Institute.

The building itself reflects the mathematics studied inside it. Among its distinctive features are mathematical patterns and structures incorporated into the architecture, including Penrose-inspired geometry, while its large central atrium creates shared spaces designed to encourage interaction across the building.

For OxML, that matters. Lectures are important, but so are the conversations between them — at poster sessions, over coffee and throughout the Institute.

It is a particularly fitting setting for several days devoted to machine learning, mathematics and scientific research.

#OxML2026 #OxML #Oxford #MathematicalInstitute #MachineLearning University of Oxford

[AI for Global Goals | OxML @GlobalGoalsAI]

AI × Mathematics — a new focus at OxML 2026

With support from Renaissance Philanthropy, OxML 2026 introduced a dedicated focus on AI for Mathematics, bringing emerging work in formalised mathematics and Artificial Mathematical Intelligence into the programme.

Two sessions approached the field from complementary directions:

🎤 Kevin Buzzard, Imperial College London — Why Formalize Mathematics
🎤 Tony Feng, UC Berkeley / Google DeepMind — The Rise and Ramifications of Artificial Mathematical Intelligence

AI × Mathematics is developing rapidly, raising fundamental questions about theorem proving, mathematical reasoning, formalisation and the role AI may play in future mathematical discovery.

This was a first step for OxML, and a direction we are keen to develop further.

#OxML2026 #OxML #AIforMathematics #ArtificialMathematicalIntelligence #MachineLearning

[AI for Global Goals | OxML @GlobalGoalsAI]

OxML happens in Oxford — and that is part of the experience.

Between lectures, poster sessions and technical discussions, OxML 2026 also gave participants time to experience the city itself.

There were walks through Oxford's historic streets and colleges, informal evenings with other participants, and plenty of conversations that continued long after the day's final session.

For an international community travelling from around the world, Oxford provides more than a backdrop. It creates a setting in which people can spend several concentrated days learning, discussing ideas and getting to know one another.

A little of OxML 2026 beyond the lecture theatre.

#OxML2026 #OxML #Oxford #MachineLearning #ResearchCommunity

[AI for Global Goals | OxML @GlobalGoalsAI]

A look back at MLx Representation Learning & Generative AI — OxML 2026

This year's track explored a broad question: how can machines learn richer representations of the world — and use them to reason, generate and act?

The programme covered intelligent data gathering, multimodal AI, embodied intelligence, multi-robot and multi-agent learning, causal discovery, AI safety and alignment, geometric deep learning, and modern approaches to generative modelling.

It also connected with this year's AI × Mathematics programme, with sessions examining both formalised mathematicsand the emerging field of Artificial Mathematical Intelligence.

A wide-ranging few days across some of the most active research directions in modern machine learning.

🎥 Explore previous OxML content on ELANDI: https://t.co/le2XNVSnaU

#OxML2026 #OxML #RepresentationLearning #GenerativeAI #MachineLearning #AI

[AI for Global Goals | OxML @GlobalGoalsAI]

A first for OxML 2026: the Poster Challenge

This year we introduced the OxML Poster Challenge, creating a dedicated space for emerging researchers to present their own work alongside the main academic programme.

Presentations spanned both Health & Bio and Representation Learning & Generative AI, with researchers discussing their work directly with participants, speakers and judges in Oxford.

We are grateful to G-Research for supporting the initiative and helping us create opportunities for researchers to participate and present their work.

The Poster Challenge was one of the most valuable additions to OxML 2026 — and something we plan to develop further for 2027.

Our amazing poster winners were:

🏆Théo Schifferli (EPFL) — CAPO: Compute-Aware Automated Protein-Model Optimization from Raw Data to Fine-Tuned Models

🏆Ben Omega Petrazzini (University of Oxford) — Transformer-based ascertainment of sudden cardiac death in electronic health records: development, validation and utility across national datasets

🏆Benjamin Frey (Helmholtz Munich) — Modelling patient variation across datasets and diseases with contrastive learning on single-cell data

🏆Nouman Ahmed (University of Oxford) — Early identification of polycystic ovary syndrome and associated multimorbidity: a multi-cohort study

🏆 Xuerui Zhang (Technical University of Munich)— Heterogeneity-Aware Distillation for Lifelong Heterogeneous Learning

🏆 Opeyemi Osakuade (The University of Edinburgh)— What Speech Tokenisers Forget: Evaluating Tone Retention in Discrete Speech Units

The challenge culminated in six $1,000 awards, recognising outstanding contributions across the two tracks.

📸 More from OxML 2026: https://t.co/zJf467THV8

#OxML2026 #OxML #PosterChallenge #MachineLearning #Research #GResearch

[AI for Global Goals | OxML @GlobalGoalsAI]

What makes OxML more than a programme of lectures?

A large part of OxML happens outside the scheduled sessions.

Over during poster discussions, breaks, between lectures and at informal socials, participants have the opportunity to speak directly with researchers, engineers and scientists working across different areas of machine learning.

For early-career researchers in particular, those conversations can matter: discussing an idea with someone whose work you know, meeting others working on the same problem, or discovering a research direction you had not previously considered.

OxML is designed to create space for those interactions alongside the formal academic programme.

Come for the research. Leave with new ideas, new questions — and, often, new people to continue the conversation with.

#OxML2026 #OxML #MachineLearning #ResearchCommunity #AI

[AI for Global Goals | OxML @GlobalGoalsAI]

A look back at MLx Health & Bio — OxML 2026

Over four days in Oxford, MLx Health & Bio brought together researchers, clinicians, engineers and scientists working at the intersection of machine learning, healthcare and biology.

The programme explored topics ranging from electronic health records, digital health and cardiovascular diagnostics to geometric deep learning, sensor-based foundation models, single-cell biology, genomics and biomedical discovery.

The academic programme was only one part of the experience. Participants also took part in:
🚶 Oxford walking tours between sessions 🤝 Socials and conversations across the community 📊 Poster sessions, where researchers presented their own work and discussed it with peers and speakers

🎥 Missed previous OxML sessions? Explore past content on ELANDI: https://t.co/le2XNVSnaU
#OxML2026 #OxML #HealthAI #MachineLearning #BioAI #AIforScience @UniofOxford @GRESEARCHjobs

[AI for Global Goals | OxML @GlobalGoalsAI]

💙 That's a wrap on MLx Representation Learning & Generative AI at OxML Oxford Machine Learning Summer School 2026! 💙

What an incredible four days of learning, collaboration and new friendships. Thank you to everyone who joined us and made this track so special - your curiosity, enthusiasm and passion for research are what make the OxML community so inspiring. A special congratulations to everyone who took part in the Poster Presentation competition. Thank you for sharing your exciting research and contributing to so many great conversations throughout the week!

To everyone who helped make this programme possible, and to every attendee who brought such fantastic energy - thank you! We hope to see many familiar faces again next year!
#OxML2026 #MachineLearning #RepresentationLearning #GenerativeAI #ArtificialIntelligence #OxfordMachineLearning #AIforGlobalGoals

[AI for Global Goals | OxML @GlobalGoalsAI]

🤖 Day 4 of MLx Representation Learning & Generative AI at OxML | Oxford Machine Learning School 2026 with Kevin Buzzard! 🤖

📣 Our speaker was Kevin Buzzard, Professor of Pure Mathematics at @imperialcollege and maintainer of Mathlib, with a talk on "Why formalize mathematics?"!

Kevin's core point was that mathematics has always split into computation (a science, with a method) and reasoning (still an art) — and that formalisation, using tools like the Lean theorem prover, is finally turning proof itself into something machine-checkable. He traced this from the slow, manual decade of building mathlib by hand, through the Polynomial Freiman-Ruzsa formalisation in 2023 that took 20 humans a month, to recent AI systems autoformalising over a million lines of Lean in weeks. His take-home: as AI starts generating mathematics faster than any human can read or trust it, formalisation stops being a nice-to-have and becomes the only way to actually know whether a machine's proof is real - checking, not producing, is where the value shifts.

👏 Thank you, Kevin!
#OxML2026 #MachineLearning #GenerativeAI #RepresentationLearning #ArtificialIntelligence #OxfordMachineLearning #AIforGlobalGoals

[AI for Global Goals | OxML @GlobalGoalsAI]

📊 Day 4 of MLx Representation Learning & Generative AI at OxML | Oxford Machine Learning School 2026 with @petergostev! 📊

📣 Our first speaker was Peter Gostev, AI Capability Lead at Arena, with a talk on "Benchmarking @ Arena"!

Peter's core argument was that no single number captures what a model can actually do — every benchmark only sees one slice of a much bigger, messier shape. He took us through how that plays out in practice: fixed test sets like GPQA and SWE-bench that get saturated and gamed, blind head-to-head comparisons on Arena that reveal real user preference instead, and newer agentic evaluations that track whether a model can actually get real work done, recover from its own mistakes, and take direction. The talk built toward a simple takeaway: we're good at measuring narrow task performance, but still bad at measuring the things that matter most for real usage — whether a model can discover something genuinely new (research), whether it holds up consistently day after day (reliability), and whether it makes sound calls in messy, real-world situations (judgment). Benchmarking, in his view, is racing to catch up with what these models are actually being asked to do.

👏 Thank you, Peter!
@GRESEARCHjobs
#OxML2026 #MachineLearning #GenerativeAI #RepresentationLearning #ArtificialIntelligence #OxfordMachineLearning #AIforGlobalGoals

[AI for Global Goals | OxML @GlobalGoalsAI]

👏 Huge congratulations to our poster presentation winners of MLx Representation Learning & Generative AI at OxML | Oxford Machine Learning School!👏

🏆 Xuerui Zhang — "Heterogeneity-Aware Distillation for Lifelong Heterogeneous Learning"
🏆 Opeyemi Osakuade — "What Speech Tokenisers Forget: Evaluating Tone Retention in Discrete Speech Units"

Very well deserved! Thank you for sharing your incredible and inspiring research! A huge thank you to Charles Martínez PhD from @GRESEARCHjobs for presenting the prizes 👏 ✨
#OxML2026 #MachineLearning #GenerativeAI #RepresentationLearning #ArtificialIntelligence #OxfordMachineLearning #AIforGlobalGoals