🏢 Google Research

User personas significantly impact the safety of large language models, bypassing safety filters more effectively than direct prompting methods.

UniAR: A unified model predicts human attention and preferences across diverse visual content (images, webpages, designs), achieving state-of-the-art performance and enabling human-centric improvement…

Transformers excel at graph reasoning, with logarithmic depth proving necessary and sufficient for parallelizable tasks; single-layer transformers solve retrieval tasks efficiently.

Learning user preferences accurately from limited data is key; this paper shows that surprisingly small datasets suffice for precise prediction, and provides efficient algorithms to achieve this.

Leveraging local simulator resets in online reinforcement learning dramatically improves sample efficiency, especially for high-dimensional problems with general function approximation.

Lowering a neural network’s geometric complexity during pre-training enhances neural collapse and improves transfer learning, especially in few-shot scenarios.

Structured Unrestricted-Rank Matrices (SURMs) revolutionize parameter-efficient fine-tuning by offering greater flexibility and accuracy than existing methods like LoRA, achieving significant gains in…

Self Logits Evolution Decoding (SLED) boosts LLM factuality by up to 20% without extra data or fine-tuning!

SequentialAttention++ unites differentiable pruning with combinatorial optimization for efficient and accurate neural network block sparsification, achieving state-of-the-art results.

Learning-based semantic routing, a scalable approach to route planning using rich user queries, is introduced, accompanied by a large-scale public benchmark and a proof-of-concept model demonstrating …

This paper reveals significant privacy gaps in shuffling-based DP-SGD, proposes a scalable Poisson subsampling method, and demonstrates its superior utility for private model training.

Randomized truthful auctions outperform deterministic ones when bidders employ learning algorithms, maximizing revenue in repeated interactions.

Prod-family algorithms achieve optimal regret in adversarial multi-armed bandits, disproving prior suboptimality conjectures.

Position coupling, a novel method, enhances the length generalization ability of arithmetic Transformers by directly embedding task structures into positional encodings. This simple technique enables…

Orchid: a novel deep learning architecture using data-dependent convolution achieves quasilinear scalability and outperforms attention-based models on various sequence modeling tasks.

Researchers created an energy-efficient optical system for generating images using light propagation, drastically reducing the latency and energy consumption of diffusion models.

MIDAS: A novel training method improves language model reasoning by efficiently stacking middle layers, surprisingly boosting downstream task performance without increasing pretraining perplexity.

MUVERA: Revolutionizing multi-vector retrieval with single-vector speed and accuracy!

Multi-turn RLHF surpasses single-turn methods by aligning LLMs with human preferences across entire conversations, not just individual turns. A novel mirror-descent algorithm, MTPO, is introduced, pr…

Linear transformers surprisingly learn intricate optimization algorithms, even surpassing baselines on noisy regression problems, showcasing their unexpected learning capabilities.