Anton van den Hengel

Frontier models don't need retraining to master lifelong embodied tasks—structuring memory handovers and spatial consolidation across fresh reasoning sessions alone drives a 22.6-point surge in multi-task navigation success.

Achieving over 90% accuracy in keypoint detection across diverse categories, GKDT sets a new standard for open-domain recognition tasks.

Hallucinated tokens in LVLMs betray themselves through diffuse attention patterns and a failure to semantically align with any specific image region, enabling highly accurate detection.

Unlock geometry-precise 3D generation by directly conditioning diffusion models on readily available point cloud priors, outperforming existing image- or text-conditioned methods.

LLMs can now explain chemical reactions from 4D molecular trajectories, thanks to a new benchmark and model designed to bridge the gap between dynamic molecular simulations and natural language understanding.