🏢 University of Oxford

LLMs generate Python heuristics for classical planning, outperforming traditional methods and challenging the state-of-the-art planning techniques.

Motion blur, usually a problem, is now a solution! This paper estimates camera motion from motion-blurred images, acting like an IMU.

VGGT: a fast, end-to-end transformer that infers complete 3D scene attributes from multiple views, outperforming optimization-based methods.

MoE-X: An intrinsically interpretable Mixture-of-Experts language model that uses sparse, wide networks to enhance transparency.

LINGOLY-TOO: A new benchmark to disentangle memorization from reasoning in LLMs using linguistic templatization and orthographic obfuscation.

OS-Genesis: Reverse task synthesis revolutionizes GUI agent training by generating high-quality trajectory data without human supervision, drastically boosting performance on challenging benchmarks.

DiTFlow: training-free video motion transfer using Diffusion Transformers, enabling realistic motion control in synthesized videos via Attention Motion Flow.

Sparse autoencoders fail to accurately decompose and interpret steering vectors due to distribution mismatch and the inability to handle negative feature projections; this paper identifies these issue…

Contrary to common belief, toxicity reduction in language models isn’t simply achieved by dampening toxic neurons; it’s a complex balancing act across multiple neuron groups.