Jun-Bo Niu

Visual grounding can now be achieved at flash speed with precise localization by treating it as parallel evidence extraction, not sequential generation.

Computer-use agents can bypass brittle, hand-engineered APIs entirely by learning when to substitute mouse clicks with native shell commands, yielding a 14.8-point leap on OSWorld.

Streaming video models consistently fail at continuous counting because they cannot distinguish re-observed entities from new instances—a failure mode resolved by framing online counting as differentiable state-machine updates over persistent visual trajectories.

Tossing out entire visual plans when robot execution drifts is needlessly wasteful: selectively branching from intermediate generative states preserves valid task structure and drives closed-loop manipulation success up to 84.8%.

TRACE achieves over 50% accuracy in long-video question answering by ensuring every answer is grounded in comprehensive visual evidence, outperforming traditional methods at a fraction of the frame cost.

LLaVA-OV-2's codec-stream tokenization lets it crush existing video-language models, especially in tasks requiring fine-grained temporal understanding of high-frequency motion.

Ditching text-based chain-of-thought unlocks better audio-visual reasoning by interleaving textual steps with a unified latent space that preserves dense sensory information.

Stop reinventing the wheel: OpenWorldLib offers a unified framework and codebase for advanced world models, finally bringing standardization to a fragmented field.

Forget bigger models: massive gains in document parsing accuracy are still possible through smarter data engineering alone.