🏢 Google DeepMind

SigLIP 2: Multilingual Vision-Language Encoders with Semantic Understanding, Localization, and Dense Features.

Eager updates drastically speed up training massive language models by cleverly overlapping communication and computation in DiLoCo, achieving near-optimal performance even with low bandwidth.

To understand AI, we need new words! This paper argues that developing neologisms—new words for human & machine concepts—is key to bridging the communication gap and achieving better AI control.

Matryoshka Quantization (MatQuant) boosts low-precision model accuracy by up to 10% through a novel multi-scale training approach. It leverages the nested structure of integer data types, allowing a …

Agency, a key concept in AI, is shown to be relative to the observer’s perspective (frame-dependent), challenging traditional binary definitions and necessitating a more nuanced approach for AI system…

AlphaGeometry2 surpasses average IMO gold medalists in solving geometry problems!

Language model distillation suffers from ’teacher hacking’, where student models over-optimize flawed teacher models, degrading true performance. This paper identifies this issue and offers effective…

AI agents now master complex tasks with improved Transformer World Models, achieving a new state-of-the-art in data-efficient reinforcement learning.

Streaming DiLoCo achieves two orders of magnitude bandwidth reduction in billion-scale parameter LLM training by synchronizing parameter subsets sequentially, overlapping communication with computatio…

TokenVerse: Extract & combine visual concepts from multiple images for creative image generation!

New multimodal safety test suite (MSTS) reveals vision-language models’ vulnerabilities and underscores the unique challenges of multimodal inputs.

Mind Evolution, a novel evolutionary search strategy, significantly boosts Large Language Model (LLM) problem-solving by generating, recombining, and refining candidate solutions via an LLM, outperfor…

Machine learning models can enable secure computations previously impossible with cryptography, achieving privacy and efficiency in Trusted Capable Model Environments (TCMEs).

Generative video models struggle to understand physics despite producing visually realistic videos; Physics-IQ benchmark reveals this critical limitation, highlighting the need for improved physical r…

Frozen LLMs get a performance boost by augmenting their key-value cache with latent embeddings generated by a differentiable offline coprocessor.

Long-context models surprisingly show that simple random sampling of examples is as effective as sophisticated methods for in-context learning, shifting the focus to efficient context utilization.

LearnLM enhances Gemini for education by training it to follow pedagogical instructions, leading to significant preference improvements over GPT-40, Claude 3.5, and Gemini 1.5 Pro in diverse learning …

PaliGemma 2: A family of versatile, open-weight VLMs achieving state-of-the-art results on various transfer tasks by scaling model size and resolution.

CAT4D: Create realistic 4D scenes from single-view videos using a novel multi-view video diffusion model.