TG-Heads

MLLMs struggle with video timing; new method recovers temporal grounding

Researchers have identified a temporal grounding issue in multimodal large language models (MLLMs) where the models understand event timing during an initial phase but lose this signal during answer generation. They dis…