Hao Huang

Muon’s unique approach to optimization reveals that sacrificing immediate gradient fidelity can significantly enhance representation utility in subsequent layers.

DualEval reveals that unifying static and preference-based evaluations can lead to more reliable model rankings and deeper insights into item performance.

SIMBA achieves superior performance in joint atmospheric profile retrieval and radiance modeling, revealing the critical importance of bidirectional processes in NWP.

Generated translation references outperform ground truth by 8.65 CEA100 points, showcasing a novel approach to literary translation challenges.