Bo Yang

VLMs still struggle to understand our planet, as revealed by a new geospatial benchmark spanning diverse Earth observation tasks and multi-source sensing data.

A new process reward model acts as a universal geospatial verifier, scaling the performance of both specialized and general-purpose VLMs in remote sensing.

Achieve state-of-the-art fine-grained vision-language alignment in remote sensing by unifying multi-granular semantic alignment and intra-modal consistency learning.

Speech, not just images, can unlock significant gains in multilingual machine translation, especially when paired with a self-evolution mechanism for MLLMs.

Forget linearization headaches: a new calculus built on naive set theory makes multiple inheritance commutative, idempotent, and associative by design.