Winner · Computer Science Subject Award 2026
Ion Stoica
“Efficient Memory Management for Large Language Model Serving with PagedAttention”
The finalists
Ten universities. Ten projects. Ten academics.
L. C. Kwek
Noisy intermediate-scale quantum algorithms
Today's quantum computers are noisy, and this review asks what can usefully be done with them anyway. Surveying the noisy intermediate-scale era, it covers the simulation of many-body physics and chemistry, combinatorial optimisation and machine learning, and argues that the period has already produced new programming paradigms that later machines will build on.
Jinlan Fu
Prompt-based learning reworks how language models are used: instead of training a model for each task, a textual prompt with unfilled slots lets a pretrained model fill in the answer. Surveying and organising research across this new paradigm in natural language processing, the work explains why the framework is so powerful and maps the emerging landscape for the field.
Sebastian Farquhar
Detecting hallucinations in large language models using semantic entropy
Large language models can reason impressively yet still hallucinate false outputs, a flaw with consequences ranging from fabricated legal precedents to risks in medical settings. The study develops entropy-based uncertainty estimators that detect a subset of hallucinations called confabulations, offering a statistical method that works even on new questions to which humans might not know the answer.
Spiros Denaxas
Prediction models in healthcare increasingly rely on artificial intelligence, and the tools used to judge their trustworthiness have had to keep pace. The project delivers PROBAST+AI, an updated framework with targeted signalling questions covering both model development and model evaluation, helping users assess the quality, risk of bias and applicability of prediction models and the studies behind them.
Ping Luo
PVT v2: Improved baselines with pyramid vision transformer
Vision transformers have driven encouraging progress in computer vision, but computational cost limits their practical use. The work improves the pyramid vision transformer with a linear complexity attention layer, overlapping patch embeddings and a convolutional feed-forward network, cutting cost to linear scale and delivering strong results on classification, detection and segmentation, with the code released openly.
Zachari Swiecki
Assessment in the age of artificial intelligence
Traditional assessment is hard to design, offers only discrete snapshots of performance, and often tests skills people now routinely use computers to perform. This paper sets out those problems, reviews artificial intelligence approaches that address them at least in part, and asks critically whether those approaches introduce fresh difficulties of their own for assessment practice.
Ion Stoica
Efficient Memory Management for Large Language Model Serving with PagedAttention
Professor Stoica directs Berkeley's Sky Computing Lab and works on cloud computing and systems for artificial intelligence. He co-authored the PagedAttention paper behind the vLLM serving engine, which manages attention key value memory efficiently to serve large language models, work carried out with collaborators at Stanford and UC San Diego.
Ghulam Muhammad
A comprehensive survey on multimodal medical signals fusion for smart healthcare systems
Professor Muhammad works on image and speech processing and intelligent healthcare at King Saud University and is a Clarivate Highly Cited Researcher. His co-authored Information Fusion survey synthesises multimodal medical signal fusion approaches for smart healthcare systems, spanning deep learning methods and clinical applications, written with collaborators at Canadian and international institutions.
Zakhar Shumaylov
AI models collapse when trained on recursively generated data
As generative models fill the internet with text, the next generation of models will inevitably train on their predecessors' output. This study shows that indiscriminate use of model-generated content in training causes irreversible defects, with the tails of the original distribution disappearing, an effect the authors call model collapse and demonstrate in language models, variational autoencoders and Gaussian mixture models.
Albert Y. Zomaya
Stochastic Client Selection for Federated Learning With Volatile Clients
Professor Zomaya directs the Centre for Distributed and High Performance Computing at Sydney. His co-authored IEEE Internet of Things Journal study tackles client selection for federated learning with volatile clients, proposing a stochastic framework that balances effective participation and fairness when devices fail unpredictably during distributed model training.
Exclusive to partners
The 2026 Subject Awards have been decided.
Partners verify their data and feature their academics. Finalists were identified from open data; winners were decided by our editorial team and announced on 15 September 2026.
Explore partnership →The wider slate
The wider slate: more candidate partnerships behind the finalists.
Towards resilience in Industry 5.0: A decentralized autonomous manufacturing paradigm
Align Your Latents: High-Resolution Video Synthesis with Latent Diffusion Models
Towards Public Verifiable and Forward-Privacy Encrypted Search by Using Blockchain
Efficient Memory Management for Large Language Model Serving with PagedAttention
Deep reinforcement learning in recommender systems: A survey and new perspectives
Transfer Learning Under High-Dimensional Generalized Linear Models
Beyond Transmitting Bits: Context, Semantics, and Task-Oriented Communications