🏢 Stanford University

ProvNeRF enhances NeRF reconstruction by modeling per-point provenance as a stochastic field, improving novel view synthesis and uncertainty estimation, particularly in sparse, unconstrained view sett…

Post-hoc model transformations can reverse performance trends, prompting a reevaluation of model selection strategies and suggesting a new ‘post-hoc selection’ method for improved model development.

Policy-Shaped Prediction (PSP) improves model-based reinforcement learning by focusing world models on task-relevant information, significantly enhancing robustness against distracting stimuli.

PaGoDA: Train high-resolution image generators efficiently by progressively growing a one-step generator from a low-resolution diffusion model. This innovative pipeline drastically cuts training cost…

Design provably convergent optimization algorithms swiftly using electric circuit analogies; a novel methodology automating discretization for diverse algorithms.

Researchers developed a verifiable training method that uses high-precision training with adaptive rounding and logging to achieve exact training replication across different GPUs, enabling efficient …

OPERA: A new algorithm intelligently blends multiple offline policy evaluation estimators for more accurate policy performance estimates.

First-Glance Off-Policy Selection (FPS) revolutionizes human-centric AI by enabling personalized policy selection for new participants without prior data, improving learning and healthcare outcomes.

OccFusion: High-fidelity human rendering from videos, even with occlusions, using 3D Gaussian splatting and 2D diffusion priors.

Newton Losses enhance training of neural networks with complex objectives by using second-order information from loss functions, achieving significant performance gains.

Scalable SEEG decoding model, seegnificant, leverages transformers to decode behavior across subjects despite electrode variability, achieving high accuracy and transfer learning capability.

Clipped SGD achieves near-optimal sub-Gaussian rates for high-dimensional heavy-tailed statistical estimation in streaming settings, improving upon existing state-of-the-art results.

MoEUT: Mixture-of-Experts Universal Transformers significantly improves the compute efficiency of Universal Transformers, making them competitive with standard Transformers in large-scale language mod…

gpSLDS, a novel model, balances expressiveness and interpretability in modeling complex neural dynamics by combining Gaussian processes with switching linear dynamical systems, improving accuracy and …

Hydra: a novel cryo-EM reconstruction method resolves both conformational and compositional heterogeneity ab initio, enabling the analysis of complex, unpurified samples with state-of-the-art accuracy…

Make-it-Real uses a large multimodal language model to automatically paint realistic materials onto 3D objects, drastically improving realism and saving developers time.

LiNGCREL, a novel algorithm, provably recovers linear causal representations from diverse environments, achieving identifiability despite intrinsic ambiguities, thus advancing causal AI.

AI agent MINIMO learns to generate challenging mathematical conjectures and prove them, bootstrapping from axioms alone and self-improving in both conjecture generation and theorem proving.

New conformal inference methods enhance LLM validity by providing adaptive validity guarantees and improving the quality of LLM outputs, addressing prior methods’ limitations.

InstructMol, a novel semi-supervised learning algorithm, leverages unlabeled data and an instructor model to significantly improve the accuracy and robustness of molecular property prediction, even wi…