🏢 Google DeepMind

Planning is redefined as a distinct inference type within a variational framework, enabling efficient approximate planning in complex environments.

LLM-powered data curation boosts web-scale visual entity recognition!

UQE: A novel query engine uses LLMs for efficient and accurate unstructured data analytics, surpassing existing methods.

Deep learning models favor simple features, hindering generalization; this paper introduces a new feature complexity metric revealing a spectrum of simple-to-complex features, their learning dynamics,…

LLMs’ inner workings remain elusive. This study uses N-gram statistics to approximate transformer predictions, revealing how LLMs learn from simple to complex statistical rules, and how model variance…

This paper presents a novel framework for estimating bounds on policy effects under unobserved confounding, offering tighter bounds and robust estimators for higher-dimensional data.

This paper introduces an innovative iterative prompting method for estimating epistemic uncertainty in LLMs, enabling reliable detection of hallucinations.

Time-reversed language models provide unsupervised feedback for improving LLMs, offering a cost-effective alternative to human feedback and enhancing LLM safety.

Finetuning’s impact on worst-group accuracy is surprisingly nuanced, with common class-balancing methods sometimes hurting performance; a novel mixture method consistently outperforms others.

Stratified Prediction-Powered Inference (StratPPI) significantly improves language model evaluation by combining human and automated ratings, using stratified sampling for enhanced accuracy and tighte…

Adaptive learning rate tuners often underperform; Curvature Dynamics Aware Tuning (CDAT) prioritizes long-term curvature stabilization, outperforming tuned constant learning rates.

Stochastic gradient bandit algorithms now guaranteed to globally converge, using ANY constant learning rate!

Simplified and generalized masked diffusion models achieve state-of-the-art results in discrete data generation, surpassing previous methods in text and image modeling.

ShiftAddLLM accelerates pretrained LLMs via post-training, multiplication-less reparameterization, achieving significant memory and energy reductions with comparable or better accuracy than existing m…

LLMs self-discover optimal reasoning structures for complex problems, boosting performance by up to 32% compared to existing methods.

Accelerate unpaired data translation with Schrödinger Bridge Flow, a novel algorithm solving optimal transport problems efficiently without repeatedly training models!

Researchers dramatically improved sign language translation by scaling up data, model size, and the number of languages, achieving state-of-the-art results.

SynCx, a novel recurrent autoencoder with complex weights, surpasses state-of-the-art models in unsupervised object discovery by iteratively refining phase relationships to achieve robust object bindi…

Researchers reveal ‘privacy backdoors,’ a new attack that exploits pre-trained models to leak user training data, highlighting critical vulnerabilities and prompting stricter model security measures.

Optimal multi-objective optimization achieved via hypervolume scalarization, offering sublinear regret bounds and outperforming existing methods.