🏢 Google DeepMind

Weak LLMs can accurately supervise strong LLMs via debate, outperforming simpler consultancy methods, especially in information-asymmetric tasks.

Normalize-and-Project (NaP) boosts reinforcement learning by stabilizing layer normalization, preventing plasticity loss, and enabling effective learning rate control.

AI models struggle with changing data; this paper introduces Soft Resets, a novel learning approach that uses an adaptive drift to gracefully guide parameters toward initialization, improving adaptabi…

Contrastive vision-language models (VLMs) trained only on English data significantly underperform on culturally diverse benchmarks. This paper reveals this bias, proposes novel evaluation metrics, and…

Neural Assets enables intuitive 3D multi-object scene editing via image diffusion models by using per-object representations to control individual object poses, achieving state-of-the-art results.

Deep learning’s mysteries surrounding sharpness regularization are solved by uncovering the crucial role of the neglected Hessian component, the Nonlinear Modeling Error (NME).

New distributional RL algorithm (DCFP) achieves near-minimax optimality for return distribution estimation in the generative model regime.

New method distills slow diffusion models into fast, few-step models by matching data expectations, achieving state-of-the-art results on ImageNet.

MooG: Self-supervised video model learns off-the-grid representations, enabling consistent scene element tracking even with motion; outperforming grid-based baselines on various vision tasks.

Data balancing in machine learning can hurt fairness and robustness; this paper reveals when and why, offering solutions for safer AI.

LLMs gain math skills via prompt-guided skill labeling and exemplar selection, significantly boosting accuracy.

Scaling up in-context learning using thousands of examples significantly boosts Large Language Model (LLM) performance, particularly for complex tasks. Novel training methods mitigate reliance on hum…

MambaLRP enhances explainability of Mamba sequence models by ensuring faithful relevance propagation, achieving state-of-the-art explanation performance, and uncovering model biases.

LLMs often generate factually inaccurate long-form text. This work introduces LongFact, a new benchmark dataset of 2280 fact-seeking prompts, and SAFE, a novel automated evaluation method that outperf…

LocCa, a novel visual pretraining paradigm, uses location-aware captioning tasks to boost downstream localization performance while maintaining holistic task capabilities.

Learn deep Successor Features (SFs) directly from pixels, efficiently and without representation collapse, using a novel, simple method combining TD and reward prediction loss!

SDF-Sim: A novel learned rigid-body simulator that leverages SDFs to achieve unprecedented scalability, enabling simulations with hundreds of objects and millions of nodes.

Gated Sparse Autoencoders (GSAEs) achieve Pareto improvement over baseline SAEs for unsupervised feature discovery in language models, resolving the shrinkage bias of L1 penalty by separating feature …

This study presents a novel Inverse Reinforcement Learning (IRL) approach for fine-tuning large language models, offering improved performance and generation diversity compared to standard methods.

Generative Hierarchical Materials Search (GenMS) uses AI to design novel crystal structures from natural language descriptions, outperforming prior methods in both fulfilling user requests and finding…