Gemini core functionality improvements
We continue to advance basic research on generative AI. In collaboration with Google DeepMind, our efforts in areas such as factuality, multilingualism, and efficiency will help improve the quality and performance of Gemini models, expand global access to our products, and better meet the needs of our users.
Our study of LLM factuality will return to our pioneering work assessing factual consistency in 2021 and conduct initial benchmarking in 2022. We will continue to advance Gemini and AI modes and publish cutting-edge research to provide fact-based information to the entire community. We published FACTS and extended it to enable robust benchmarking of facticity in LLM and techniques that improve factivity, such as text-to-image conversion, video generation, long context, and uncertainty representation.
At I/O, we’ve seen that moving information is becoming more complex and people are having longer conversations to get what they need. This poses several challenges for LLMs, including being able to infer and analyze more relevant information in the context window, following constraints that emerge earlier in the conversation, and using longer reinforcement learning trajectories. Google Research is pioneering efforts on all of these challenges, and these advances drive our Gemini model.
You can also ask complex and long questions in Google Maps using the new Ask Maps feature. We partnered with Ask Maps to upgrade its evaluation framework and redefine how map usefulness is measured. This collaboration established a critical feedback loop that is essential to the continuous improvement of Ask Maps’ performance by accurately identifying complex edge cases involving model inference and tool execution. We also drove research to improve the quality of Ask YouTube, a new feature that helps users easily find videos and information.
Generative AI will make tools and products much more accessible, allowing technology to finally be available where users are. We’ve been evolving Gemini’s multilingualism and localization capabilities, including publishing benchmarks that show how LLM works in different languages and different locations, and open sourcing data for African languages developed with the community. Our efforts have enabled us to expand Gemini to over 230 countries and over 70 languages. This makes Gemini the world’s most widely available AI assistant.
Google builds low-latency, high-throughput infrastructure to meet the needs of users, developers, and businesses around the world. Our research team has developed new techniques built on speculative decoding, including block validation and tree-structure drafting. This involves intelligently exploring multiple continuation candidates at once, accepting more tokens at each step. Our implementation is highly optimized for Google’s TPU architecture, maximizing hardware utilization and providing significantly faster response without compromising quality. This effort led to the current speed of Gemini 3.5 Flash, and the same model is now available in Antigravity and AI Studio.
