Siqi Luo

A unified framework reveals that existing LLM policy optimization methods often overlook compound failures that require simultaneous adjustments to both trajectory and reward components.

Even state-of-the-art T2I models falter in generating scientifically accurate illustrations, revealing critical gaps in text-rendering and reasoning capabilities.

LLMs are still far from being able to generate expert-level clinical guidelines, despite advances in deep research systems.