Irina Nikishina

Most web-browsing LLM agents struggle to find obscure, multi-hop information across diverse modalities and languages, even when explicitly designed for web interaction.

Idiom comprehension in low-resource languages suffers significantly, with literal meanings proving far more challenging than figurative interpretations, even in context-rich conversations.

LLMs can fact-check surprisingly well without external retrieval, and the key is tapping into their internal representations, not just the output logits.