🏢 University of Oxford

AI agents learn better with well-designed training environments. This paper reveals flaws in current environment-selection methods and introduces Sampling for Learnability (SFL), a new approach that …

Multimodal models’ impressive ‘zero-shot’ performance hinges on the frequency of concepts in their training data, not inherent generalization ability; exponentially more data is needed for linear impr…

METRIC FLOW MATCHING (MFM) generates smooth interpolations on data manifolds by minimizing kinetic energy, outperforming Euclidean methods and achieving state-of-the-art results in single-cell traject…

Frugal Flows: Generate realistic causal benchmarks with exact marginal causal effects, enabling robust causal method validation.

LoCo: Memory-efficient location-consistent image features learned via a novel ranking loss, enabling three orders of magnitude memory improvement and outperforming state-of-the-art.

This paper introduces a novel unsupervised video object segmentation method using long-term point trajectories and optical flow, outperforming prior art by effectively combining sparse, long-term moti…

Learn on massive graphs efficiently using Intersecting Community Graphs (ICGs)! This method approximates large graphs with ICGs, enabling linear time/memory complexity for node classification.

Language models struggle with hierarchical information. This work introduces Hierarchy Transformer Encoders (HITs), a novel method to retrain transformer encoders using hyperbolic geometry and special…

Bridging the accuracy gap in online continual learning caused by label delays, a new framework with Importance Weighted Memory Sampling prioritizes relevant memory samples, significantly outperforming…

This paper introduces a novel framework for creating interventionally consistent surrogate models for complex simulations, addressing computational limitations and ensuring accurate policy evaluation.

Researchers developed methods to measure and interpret the divergence between learned feedback patterns (LFPs) in LLMs and human preferences, helping minimize discrepancies between LLM behavior and tr…

New modeling of bid space in multi-unit uniform price auctions achieves regret of Õ(K4/3T2/3) under bandit feedback, improving over prior work and closing the gap with discriminatory pricing.

G2D: a novel medical VLP framework achieves superior performance in medical image analysis by simultaneously learning global and dense visual features using image-text pairs without extra annotations.

FilterNet: A novel deep learning architecture using learnable frequency filters for superior time series forecasting accuracy and efficiency.

Small ensembles in stochastic linear bandits achieve near-optimal regret; a rigorous analysis shows that ensemble size need only scale logarithmically with horizon.

Direct3D: Revolutionizing image-to-3D generation with a scalable, native 3D diffusion model achieving state-of-the-art quality.

BAL-PM, a novel active learning approach, drastically reduces human feedback in LLM preference modeling by leveraging both model uncertainty and prompt distribution diversity, achieving 33%-68% fewer …

COUNTGD: A new multi-modal model counts objects in images using text or visual examples, significantly improving open-world counting accuracy.

Learned optimizer OPEN tackles RL’s non-stationarity, plasticity loss, and exploration using meta-learning, significantly outperforming traditional and other learned optimizers.

LLM agents surprisingly exhibit human-like trust behavior, especially GPT-4, paving the way for simulating complex human interactions in various applications.