Improving core abilities in Gemini
We continue to pursue fundamental research for Generative AI. In collaboration with Google DeepMind, our work in the areas of factuality, multilingualism and efficiency helps advance the quality and performance of Gemini models and expand the global reach of our products to better meet users’ needs.
Our research on LLM Factuality in 2021 is based on leading research on the assessment of factual consistency and an initial benchmark In 2022. We will continue to advance Gemini and AI Mode, and publish cutting-edge research to help provide factual information to the entire community. we have published fact And increased it LLM includes techniques to allow robust benchmarking of factuality, and to improve factuality. text-to-image, video generation, long-term context And expression of uncertainty.
At I/O, we saw that information journeys are becoming increasingly complex, where people engage in longer conversations to get what they need. This creates several challenges for LLMs, including being able to reason and analyze more relevant information in the context window, observing constraints that appear early in the interaction, and using longer reinforcement learning trajectories. Google Research has done pioneering work on all of these challenges, and these advances fuel our Gemini model.
New ask map The feature also allows people to ask complex, lengthy questions in Google Maps. We partnered with Ask Maps to advance its evaluation framework and redefine how we measure map usefulness. By pinpointing complex edge cases involving model logic and tool execution, this collaboration established a critical feedback loop – critical to continuously improving the performance of Ask Maps. We also conducted research to improve quality ask youtubeA new feature that helps users find videos and information easier.
Generative AI is making tools and products more accessible, and allowing technologies to finally meet users where they are. We’ve advanced multilingualism and localization capabilities for Gemini, including the publication of a benchmark that shows how LLMs work in different languages, and various places, And Open sourcing data in African languagesEvolved with the community. Our efforts helped Gemini expand to more than 230 languages in more than 70 countries. This makes Gemini the most widely available AI assistant in the world.
Google builds its infrastructure to achieve low latency and high throughput, so we can meet the needs of users, developers, and enterprises around the world. Our research teams developed new techniques including speculative decoding block verification and tree-structured drafting, which intelligently detects multiple candidate continuations simultaneously and accepts more tokens per step. Our implementation is highly optimized for Google’s TPU architecture, which maximizes hardware utilization to deliver significantly faster responses without any loss in quality. This work enabled the current pace gemini 3.5 flashThe same models also power AntiGravity and AI Studio.
