🏢 University of Oxford

Safety fine-tuning for LLMs is shown to minimally transform weights, clustering inputs based on safety, but is easily bypassed by adversarial attacks.

First unsupervised object detection method with theoretical guarantees to recover true object positions, up to quantifiable small shifts!

Fully recurrent neural networks can be universal in-context approximators, achieving the same capabilities as transformer models by cleverly using prompts.

Large language models (LLMs) suffer from information loss due to representational collapse and over-squashing, causing failures in simple tasks; this paper provides theoretical analysis and practical …

GroupEXP-DA boosts domain adaptive 3D object detection by using a grouping-exploration strategy to reduce bias in pseudo-label collection and account for multiple factors affecting object perception i…

Offline model-based RL methods fail as dynamics models improve; this paper reveals the ’edge-of-reach’ problem causing this and introduces RAVL, a simple solution ensuring robust performance.

D³GM, a novel score-based diffusion model, enhances stability & generalizability in solving inverse problems by leveraging measure-preserving dynamics, enabling robust image reconstruction across dive…

SpatialPIN boosts vision-language models’ spatial reasoning by cleverly combining prompting techniques with 3D foundation models, achieving zero-shot performance on various spatial tasks.

Set-based Neural Network Encoder (SNE) efficiently encodes neural network weights for property prediction, eliminating the need for architecture-specific models and improving generalization across dat…

Sequence-augmented SE(3)-Flow model, FOLDFLOW-2, excels at generating diverse, designable protein structures, surpassing existing methods in unconditional and conditional design tasks.

Transformers and RNNs show contrasting representational capabilities: Transformers excel at tasks requiring associative recall, while RNNs are better suited for hierarchical language processing. This …

Researchers developed a novel algorithm to automatically find optimal schedules for denoising diffusion models (DDMs), significantly improving sample quality and efficiency without manual parameter tu…

Rough Transformers: A lightweight continuous-time sequence modeling approach using path signatures to significantly reduce computational costs, improving efficiency and accuracy, particularly for long…

Random pixel projections outperform complex online continual learning methods for image classification, challenging assumptions about representation learning.

COBOL: a novel Bayesian Optimization algorithm leverages human expert advice via binary labels, achieving both fast convergence and robustness to noisy input, while guaranteeing minimal expert effort.

Pre-trained text-to-image diffusion models create highly effective, versatile representations for embodied AI control, surpassing previous methods.

OxonFair: a new open-source toolkit for enforcing fairness in binary classification, supporting NLP, Computer Vision, and tabular data, optimizing any fairness metric, and minimizing performance degra…

This research introduces Cluster-guided Sparse Experts (CSE), enabling pretrained language models to effectively learn long-tail domain knowledge without domain-specific pretraining, thus achieving su…

Fractal dimension, while showing promise, fails to consistently predict neural network generalization due to hyperparameter influence and adversarial initializations; prompting further research.

Extrapolated FTRL ensures Nash equilibrium convergence in harmonic games, defying standard no-regret learning limitations.