Mixture of Experts (MoEs) offer a sparse alternative to dense language models by routing inputs to only a subset of expert networks, reducing computational cost and memory usage while maintaining performance. This approach addresses the practical limitations of scaling dense models, such as high training costs and inference latency. Hugging Face has enhanced its Transformers library with tools like Dynamic Weight Loading and Lazy Materialization to support efficient MoE implementation.

Circle to Search has been updated to allow users to search for multiple objects within a single image simultaneously by circling them and asking a question.

ImpRIF is a method that improves large language models’ ability to follow complex instructions by enhancing their understanding of implicit reasoning through verifiable reasoning graphs and graph-driven chain-of-thought reasoning. The approach synthesizes training data, uses fine-tuning and reinforcement learning, and achieves significant performance gains on multiple benchmarks.

Explore-on-Graph (EoG) introduces a reinforcement learning framework that incentivizes large language models to autonomously explore diverse reasoning paths on knowledge graphs, using path-refined reward modeling to improve accuracy and reduce hallucinations. The method outperforms both open-source and closed-source LLMs on multiple KGQA benchmarks by enabling more flexible and generalizable reasoning beyond predefined patterns.

Google has released Gemini 3, its most advanced foundation model yet, achieving record-breaking scores on reasoning benchmarks like Humanity’s Last Exam and now available through its app and AI search. The model is positioned as a top contender in the rapidly evolving AI market, with a deeper version, Gemini 3 Deepthink, upcoming for premium subscribers after safety testing.

Samsung Galaxy S26 users will receive the latest Google AI features through updated Android, transforming the OS into a more intelligent, learning-driven system that adapts to user behavior.

OpenAI and Figma have launched a Codex integration that links code and design, allowing developers and designers to switch seamlessly between implementing code and editing designs in Figma for faster iteration and development.

This paper presents a systematic study of scaling laws for view synthesis transformers, showing that encoder-decoder architectures can be compute-optimal and outperform decoder-only models with lower training compute while achieving state-of-the-art performance in novel view synthesis.

This paper proposes using multi-way parallel text alignment to improve multilingual embeddings by training on diverse language translations, resulting in better cross-lingual alignment and performance across NLU tasks and benchmarks like MTEB. It demonstrates significant gains in bitext mining and model fine-tuning, even with small datasets, highlighting the value of multi-way supervision.

OpenAI has appointed Arvind KC as Chief People Officer to support company scaling, cultural development, and adapting workplace practices in the era of AI.