Xin Wei

Exploiting local parameter perturbations of a teacher model unlocks complementary supervision signals, boosting LLM mathematical reasoning accuracy beyond standard on-policy self-distillation.

MLLMs can now seamlessly unify perception and reasoning, achieving state-of-the-art performance in free-form multimodal grounding tasks.

GeoAnchor achieves superior 3D spatial reasoning by integrating multiple latent representations, outperforming existing methods that struggle with complex geometric scenarios.

Static scanners fail against adaptive evasions, but a new behavior-centric auditor can detect 97% of malicious skills with minimal false positives.