Michael Färber

Tactile representations in robots struggle to generalize, with performance dropping to just 25.1% Recall@1 when faced with unseen materials.

Language models struggle with cultural nuances, but NRITYAM reveals significant gaps in their understanding of global dance traditions.

Machine-translated benchmarks can be unreliable, but this automated pipeline offers a practical, scalable way to flag low-quality translations and prioritize human review.

Surprisingly, even with carefully designed retrieval pipelines, smaller language models still fall short of larger models on complex scientific reasoning tasks, suggesting that retrieval and model scale are complementary, not interchangeable.