Adib Sakhawat

A comprehensive dataset of Bangladeshi legal acts reveals over two centuries of legislative evolution, enabling unprecedented multilingual and historical analysis in legal NLP.

Multilingual embeddings' cross-lingual retrieval failures aren't due to anisotropy as commonly believed, but rather to "hubness," a previously underappreciated geometric pathology that can be largely fixed with a simple score correction.

LLMs are surprisingly bad at strategic deduction in multi-turn dialogues, despite excelling at information containment, revealing a critical asymmetry in their reasoning abilities.

LLMs can be surprisingly resistant to persuasion even when they seem persuasive, revealing a fundamental asymmetry in their social intelligence.