Compares LoRA versus full-parameter finetuning on challenging code and math tasks; sheds light on the learning-forgetting tradeoffs. Showing that LoRA usually underperforms full finetuning in a new target domain while forgetting less of the source domain.

Small on-device LMs guided by frontier cloud-hosted LMs to solve workloads on device at the fraction of the cost.

Introduces a semi-supervised approach to pose estimation, using physically-informed inductive biases to improve generalization with fewer labels. Poses are further refined by combining deep ensembles with state-space models. Open-sourcing a deep learning system that is optimized for efficiency, building on PyTorch Lightning and NVIDIA DALI.

This model disentangles movement that can be quantified by keypoints (e.g., limb position) from subtler feature variations like orofacial movements. We introduce a novel VAE whose latent space decomposes into two orthogonal subspaces – one unsupervised subspace and one supervised subspace linearly predictive of labels (keypoints). The latent space additionally includes a context variable that predicts the video/subject identity.

Deriving systematic biases in scalable GPs; eliminating them using Russian-Roulette estimators.

Estimate pose using a CNN; Model pose as a rigid-body system; infer its angular dynamics AND the external torques acting on it using a new nested variant of VSMC.