Kyomin Jung

Language-dependent asymmetries in instruction hierarchy compliance reveal that multilingual models can be more reliable in cross-language contexts, but risk security vulnerabilities due to language specialization.

Existing diversity metrics miss the mark, failing to capture the true variety in problem-solving strategies that could enhance LLM reasoning capabilities.

LVLMs can generate more factual and detailed image captions at a lower compute cost by reflecting on their past mistakes and systematically attending to overlooked details.