Subject Award · Physical Sciences & Engineering · 2026
The ten finalist universities are the ten highest-ranked on the Computer Science collaboration index, listed here alphabetically until the full index publishes on 15 September. Each finalist is represented by an academic whose recent work exemplifies why: a real project, drawn from the open scholarly record, cited so you can check it.
The finalists
Body-Worn Sensors for Recognizing Physical Sports Activities in Exergaming via Deep Learning Model
Physical games that get people moving are the starting point here, with body-worn sensors recognising sports activities through a deep learning model. Beyond fitness, the team present the system as multi-purpose: trained on a domain-specific dataset, the same gesture recognition and virtual reality depiction pipeline is proposed for other application areas.
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.
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.
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.
Federated Learning for Smart Healthcare: A Survey
Hospitals hold data that could improve care, but pooling it centrally is often impractical and raises real privacy concerns. This survey examines federated learning for smart healthcare, in which multiple institutions coordinate to train shared artificial intelligence models without exchanging raw data, covering the motivations, the requirements and the recent designs proposed for the field.
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.
Parameter-efficient fine-tuning of large-scale pre-trained language models
As language models grow, retraining and storing every parameter becomes prohibitively costly. This work reviews parameter-efficient adaptation, where only a small portion of a model's parameters is optimised while the rest stay fixed, and gathers the various designs under the single term 'delta-tuning', showing that very large models can be stimulated effectively by tuning very little.
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.
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.
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.
Exclusive to partners
Partners verify their data, feature their academics, and are eligible for the Subject Awards. Finalists are identified from open data; winners are decided by our editorial team and announced on 15 September 2026.
Explore partnership →The wider slate
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