Jaewoo Kang

GUI agents forget their way: dynamic memory modality selection unlocks significant performance gains over static approaches.

Multimodal models frequently bypass their supposed latent reasoning chains via lingering visual caches—severing these bypass paths proves that recurrence can genuinely carry the full inferential burden without sacrificing visual competence.

Activating task-relevant latent features from high-resource languages can significantly boost reasoning performance in low-resource languages, challenging assumptions about language capabilities.