Xue-Liang Zhao

Frontier MLLMs exhibit strong local perception but falter on long-horizon audio-visual reasoning, with a single benchmark exposing specific failure modes like identity drift and cross-modal misalignment.

EntroRouter achieves near-optimal expert accuracy while slashing computational costs by nearly half, challenging the status quo of model routing efficiency.

Training with large block sizes cripples reasoning performance, but a novel curriculum approach unlocks strong reasoning capabilities in diffusion models.

A self-evolving training framework that tailors spatial reasoning tasks to a model's current capabilities can dramatically enhance performance while minimizing data inefficiency.

Forget left-to-right: Dream-Coder 7B's diffusion approach lets it generate code in *any* order, adapting its strategy to the task at hand.