Xilin Wang

Most alignment model organisms degrade core capabilities and naturalness during training, undermining interpretability findings; a new multi-objective approach preserves these crucial aspects.

Lifelong navigation agents collapse when environments change over time, exposing a critical blind spot in how embodied models cache and reuse prior spatial memories.

Outperforming previous methods, UniLM-Nav achieves zero-shot last-mile navigation by effectively integrating multimodal reasoning and task context.