Hongbo Wang

Conservation laws can survive learned representations, but their certifiability varies dramatically depending on the underlying geometric structure.

Trust-horizon certification achieves zero violations in 50 audits, revealing that equivariant models can maintain robust performance across diverse environments.

Equivariant models can provide a certified predictable horizon, while non-equivariant models are fundamentally limited in their predictive reliability.

LAFP achieves up to 15% higher success rates in imitation learning by preserving the multimodal structure of latent actions, challenging the limitations of traditional behavior cloning.

Hallucinated details in super-resolution are not just random noise; they reveal a fundamental spectral mismatch that can be corrected by shaping the generative process itself.

Chinese toxicity detectors are surprisingly easy to fool with subtle semantic indirection and obfuscation, missing almost 70% of attacks generated by the CITA framework.

LLMs struggle to navigate the complex, multi-turn justification and response dynamics of real-world patent examination, revealing critical gaps in legal reasoning and technical novelty judgment.