We reproduce and compare different distillation strategies for cross-encoders across a wide range of architectures, identifying robust design choices for effective re-ranking.

We propose MICE (Minimal Interaction Cross-Encoders), a new architecture that achieves a 4x decrease in inference latency while maintaining high ranking effectiveness.

We introduce ToMMeR, a lightweight model probing mention detection capabilities from early LLM layers, achieving high zero-shot recall across diverse benchmarks.

We investigate how Large Language Models internally represent entities by introducing entity mention reconstruction as a novel framework for studying how LLMs encode and manipulate entities.

We study the numerical approximation of scalar conservation laws by computational optimization of the numerical flux function in a first-order finite volume method.