Search papers, labs, and topics across Lattice.
3
0
6
AI-generated image detectors can be misled by spurious features, but DEAR prunes these distractions to enhance detection robustness against post-processing.
Sparse autoencoder features in CLIP aren't just about *what* they represent, but *where* they look, and that "information scope" fundamentally shapes how they influence CLIP's decisions.
Forget siloed robots: MLLMs can orchestrate sensing, communication, and computation across robot networks to achieve superhuman task performance.
