Search papers, labs, and topics across Lattice.
University of Toronto
6
1
8
22
Disentangled representations can reduce unlearning collateral damage by up to 4x, challenging the assumption that entanglement is merely a theoretical concern.
Minimizing distortion, not maximizing codebook utilization, is the key to achieving superior reconstruction fidelity in visual tokenization.
VQ-Transplant slashes training costs by 95% while maintaining near state-of-the-art reconstruction quality, making advanced VQ techniques accessible to resource-constrained researchers.
Insertion Language Models can achieve competitive performance with left-to-right and masked diffusion models while offering enhanced sampling flexibility through a novel continuous-time Markov chain framework.
Turns out, a single, well-tuned temperature can drastically improve the reliability of language model confidence scores, outperforming more complex recalibration schemes.
Straightening latent space trajectories with a simple curvature regularizer dramatically improves the stability and success of gradient-based planning in world models.
