Abstract: We conduct a systematic analysis that uncovers how visual perception and message‑passing offer complementary strengths in graph understanding, and introduce a novel benchmark to showcase these insights. Our findings reveal that vision models can significantly enhance graph structural understanding, outperforming traditional GNNs in various tasks.

Abstract: We propose Political-LLM, a framework that bridges large language models with political science. It provides a dual-perspective taxonomy, political tasks and computational methods, while outlining key challenges and future directions, aiming to guide ethical and effective AI use in political research.

Abstract: We present MJ-VIDEO, a Mixture-of-Experts reward model for fine-grained video preference evaluation, which is built upon MJ-BENCH-VIDEO, a large-scale benchmark covering alignment, safety, coherence, and bias. Our model achieves significant improvements in preference judgment and enhances alignment in video generation.