Yue Cheng

RLVR's reasoning boost isn't just about more data – it's a Goldilocks problem: too easy and LLMs skip the reasoning, too hard and they break down, but just right and they become reasoning powerhouses.

Reconstructing 3D facial features from 2D images significantly boosts deepfake detection, achieving state-of-the-art performance and robustness.

Hallucinating LLMs in enterprise workflows can be tamed by a new Hybrid Utility Minimum Bayes Risk (HUMBR) framework that synthesizes semantic and lexical signals to achieve consensus without ground truth.