Katharina von der Wense

Monolingual RAG performance significantly improves when the LLM's internal reasoning language is aligned with the query and retrieved documents, even if the model is more proficient in another language.

Multilingual medical benchmarks penalize cross-lingual answer variation as model error, but real-world clinicians are sharply split on whether AI should enforce universal consistency or adapt to local cultural contexts.

MLLM-generated text shows distinct signs of translationese, but surprisingly, it diverges from traditional translation patterns in significant ways.

LLMs' internal confidence signals may mislead, as our verbalized methods reveal a stark disconnect between certainty and correctness in machine translation.

Instruction-tuned LLMs can be up to 26% overconfident in their own responses, but a simple adjustment can significantly improve their calibration.

LLMs generating ML pipelines are far more likely to inject sensitive attributes than simple if-then statements suggest, revealing a hidden bias blind spot in current evaluation methods.

LLMs fail spectacularly at understanding and generating Meenzerisch, a German dialect, achieving less than 10% accuracy even with targeted interventions, revealing a significant gap in their linguistic capabilities for low-resource languages.