Yu Huang

Most video generation models fail to render text accurately, with even top performers exhibiting a 25% word error rate on a new benchmark designed for this critical oversight.

Even the strongest LLM agent workflows top out at just 25.4% proof coverage on STOC and COLT theorems, exposing a massive capability overhang between competition math performance and genuine theoretical computer science research.

Qualitative examples show the PhysBrain 1.5 model's ability to produce end-effector trajectories and predict future scenes through spatially aligned RGB, depth, and robot-mask outputs.

OPD may improve sampling efficiency, but it risks making previously solvable problems unsolvable, challenging the notion of true capability expansion in LLMs.

GraphAlignCoder boosts code generation accuracy by 31.6% to 43.8% by embedding formal proof structures into the training process.

Knowledge hallucinations are the most significant hurdle for MLLMs, revealing critical weaknesses in their robustness to false premises.

Novice developers using AgentForge significantly improved their software engineering skills and critical collaboration with AI, despite facing varying interaction challenges.

Specula uncovers deep bugs in system code that traditional methods often miss, revolutionizing formal specification generation.

GUI agents struggle with long tasks not because they mis-click, but because they forget what they were doing, and a new "anchored memory" method can fix it.

Shrinking visual document retrieval storage by 95% is now possible without sacrificing accuracy, thanks to a layout-aware parsing strategy.

The first comprehensive survey of Visual Document Retrieval reveals how MLLMs are reshaping the field, highlighting the shift towards RAG and agentic systems for complex document understanding.