🏢 ETH Zurich

WildGaussians enhances 3D Gaussian splatting for real-time rendering of photorealistic 3D scenes from in-the-wild images featuring occlusions and appearance changes.

TACOS: A novel time-adaptive RL framework drastically reduces interactions in continuous-time systems while improving performance, offering both model-free and model-based algorithms.

Weight decay in deep learning surprisingly induces low-rank attention layers, potentially harming performance but offering optimization strategies for large language models.

MMVM VAE enhances multimodal data analysis by using a soft constraint to guide each modality’s latent representation toward a shared aggregate, improving latent representation learning and missing dat…

UniSDF: Unifying neural representations reconstructs complex scenes with reflections, achieving state-of-the-art performance by blending camera and reflected view radiance fields.

UniBias unveils and mitigates LLM bias by identifying and eliminating biased internal components (FFN vectors and attention heads), significantly improving in-context learning performance and robustne…

Unifying framework reveals hidden connections between attention, recurrent, and state-space models, boosting foundation model efficiency.

New methods minimize outlier features in transformer training, improving quantization and efficiency without sacrificing convergence speed.

This paper introduces transductive active learning, proving its efficiency in minimizing uncertainty and achieving state-of-the-art results in neural network fine-tuning and safe Bayesian optimization…

Testably learning polynomial threshold functions efficiently, matching agnostic learning’s best guarantees, is achieved, solving a key problem in robust machine learning.

SWT-Bench, a new benchmark, reveals that LLMs excel at generating tests for real-world bug fixes, surpassing dedicated test generation systems and significantly improving code-fix precision.

Neural network hyperparameter transferability across vastly different model sizes is achieved via a newly discovered property called ‘Super Consistency’ of loss landscapes.

Stochastic Concept Bottleneck Models (SCBMs) revolutionize interpretable ML by efficiently modeling concept dependencies, drastically improving intervention effectiveness and enabling CLIP-based conce…

SPEAR, a novel algorithm, precisely reconstructs entire data batches from gradients in federated learning, defying previous limitations and enhancing privacy risk assessment.

TVSAFEOPT: Safe time-varying optimization using spatio-temporal kernels ensures safety while tracking time-varying reward and safety functions, providing optimality guarantees in stationary settings.

Outlier-robust mixture learning gets order-optimal error guarantees, even when outliers massively outnumber small groups, via a novel meta-algorithm leveraging mixture structure.

Recurrent neural networks struggle with long-term memory due to a newly identified ‘curse of memory’: increasing parameter sensitivity with longer memory. This work provides insights into RNN optimiza…

QuaRot: Revolutionizing 4-bit LLM inference with lossless quantization via rotation!

This paper delivers a groundbreaking polynomial-time algorithm for optimally estimating edge density in random graphs while ensuring node privacy and robustness against data corruption.

POSEIDON: a novel foundation model for PDEs achieves significant gains in accuracy and sample efficiency, generalizing well to unseen physics.