Siyuan Li

Transition-level supervision can dramatically enhance multimodal model performance, revealing that coherence between text and visuals is crucial for complex reasoning tasks.

LLMs struggle with complex legal rules, often missing critical exceptions, but a new benchmark and structured representation can dramatically enhance their compliance accuracy.

Zero-shot learning can now predict traffic accidents in real-time without the need for costly annotated datasets, achieving competitive results in a major competition.

PaperFlow redefines scientific paper recommendation by adapting to user interests over time, achieving unmatched alignment with real-world reading behaviors.

Freezing your VQ decoder during text-to-image post-training might be why your images are getting worse even as your CLIP scores improve.

AI research agents can now reliably trace method evolution topologies thanks to a new methodological evolution graph, Intern-Atlas, that captures structured relationships between research methods.

A principled framework for General World Models reveals the limitations of current systems and the architectural requirements for future progress.