🏢 University of Cambridge

Zero-Shot Tokenizer Transfer (ZeTT) detaches language models from their tokenizers via a hypernetwork, enabling efficient on-the-fly tokenizer swapping without retraining, significantly improving LLM …

Zero-shot RL struggles with low-quality data; this paper introduces conservative algorithms that significantly boost performance on such data without sacrificing performance on high-quality data.

TinyTTA enables efficient test-time adaptation on memory-constrained edge devices using a novel self-ensemble and early-exit strategy, improving accuracy and reducing memory usage.

TabEBM: Class-specific EBMs boost tabular data augmentation, improving classification accuracy, especially on small datasets, by generating high-quality synthetic data.

Self-healing machine learning (SHML) autonomously diagnoses and fixes model performance degradation caused by data shifts, outperforming reason-agnostic methods.

SOFO: a novel second-order optimizer enables efficient and memory-friendly RNN training for neuroscience tasks, surpassing Adam’s performance, especially on long time horizons.

LLMs struggle with out-of-distribution (OOD) generalization. This research introduces ‘rule extrapolation’ using formal languages to rigorously evaluate OOD behavior in various LLM architectures, rev…

This research unveils a compute-optimal recipe for fine-tuning language models into high-quality text embedding models, offering practical guidance and scaling laws for resource-constrained settings.

Relational Concept Bottleneck Models (R-CBMs) merge interpretable CBMs with powerful GNNs for high-performing, explainable relational deep learning.

CordsNet: a hybrid CNN-RNN architecture enabling biologically realistic, robust image recognition through continuous-time recurrent dynamics.

Deep learning algorithms now predict quantum ground state properties with constant sample complexity, regardless of system size, improving upon previous methods.

Predicting RL agent behavior is key for safety and interaction; this study reveals that explicitly planned agents are significantly easier to predict due to their internal plans.

µPOCA: a new active learning approach maximizes model generalization using strategically acquired labels/features in data-scarce, costly scenarios with partial observability, leveraging LLMs for effic…

This paper introduces novel autoregressive sampling and hybrid training strategies for score-based diffusion models, significantly boosting PDE forecasting and assimilation accuracy.

Researchers introduce Geometric Parameterization (GmP), a novel neural network parameterization resolving instability in ReLU network training, leading to faster convergence and better generalization.

Researchers created MMA, a large multilingual dataset of informal-formal mathematical pairs, leveraging a language model for reverse translation. Fine-tuned models achieved significantly improved aut…

Localized Adaptive Risk Control (L-ARC) improves fairness and reliability of online prediction by providing localized statistical risk guarantees, surpassing existing methods in high-stakes applicatio…

Accelerate Gaussian process hyperparameter optimization by up to 72x using novel linear system solver techniques.

HEALNet: a novel multimodal fusion network achieving state-of-the-art performance on biomedical survival analysis by effectively integrating heterogeneous data while handling missing modalities.

GRANOLA: A novel graph-adaptive normalization layer significantly boosts GNN performance by dynamically adjusting node features based on the input graph’s unique structure.