B. Schiele

Discovering task-specific, compact concept vocabularies from unsupervised pools unlocks inspectable AI without the need for manual annotations or massive, static concept sets.

Data mixing, especially with instruction-heavy data, emerges as the crucial factor for optimizing VLM training, challenging traditional filtering approaches.

PARCEL redefines visual tokenization, achieving superior efficiency and performance by dynamically anchoring feature extraction to spatial pool tokens.

Seemingly beneficial architectural choices for end-to-end driving, like high-resolution perception, can actually hinder scalable closed-loop performance, highlighting the need for careful co-design.

Mechanistic interpretability gets a formal footing: "Certified Circuits" uses data subsampling to find provably stable sub-networks, boosting accuracy by up to 91% while using 45% fewer neurons.