Google Research at Google I/O 2025

Each year at Google I/O, we share some of Google’s most advanced technologies. We show how they can be helpful and provide new experiences, and how developers and other communities can use them to innovate. Many of these new technologies emerged from years of work within Google Research, many in collaboration with other teams, building on multiple successive breakthroughs in AI and other areas of computer science. This year’s I/O highlights the impact of bringing research to reality. As Sundar put it: “What all this progress means is that we’re in a new phase of the AI platform shift. Where decades of research are now becoming reality for people, businesses and communities all over the world.”

In addition to Google Research’s contributions to Gemini and the generative AI products highlighted on the I/O stage, here are some of our favorites this year, tapping into years-long efforts from Google Research to realize the magic cycle of research.

Since we first introduced Med-PaLM in 2022, followed by Med-PaLM2 and Med-Gemini, our research teams have been continuously advancing AI to make healthcare more accessible and effective. At I/O, we announced MedGemma, Google’s most capable open model for multimodal medical text and image comprehension. It has the potential to speed up the development of new healthcare products.

MedGemma is based on Gemma 3 and designed to be a starting point for developers building health applications, such as analyzing radiology images or summarizing clinical data. Its small size makes it efficient for fine-tuning for specific needs, and when evaluated on the MedQA benchmark, its baseline performance on clinical knowledge and reasoning tasks is similar to that of much larger models. Since MedGemma is open, it can be run in the developer’s preferred environment, including on the Google Cloud Platform or locally. Both MedGemma 4B and the 27B text-only models are now available on HuggingFace and Vertex Model Garden as part of our Health AI Developer Foundations (HAI-DEF).

MedGemma follows our recent announcement about AMIE, developed in collaboration with Google DeepMind, that was also highlighted at I/O. AMIE is a research AI agent for medical diagnostic conversations. The new multimodal version can intelligently interpret and reason about visual medical information, helping clinicians towards more accurate diagnoses.

For nearly two years, our teams at Google Research and across Google have been collaborating with educational experts on LearnLM, a family of fine-tuned models for learning. At I/O we announced that LearnLM will now be available directly in Gemini 2.5, making it the world’s leading model for learning. Our latest tech report demonstrates that Gemini 2.5 Pro outperforms alternative models on learning science principles and is the preferred choice for educators. It has advanced STEM reasoning, multimodal understanding, quizzing and assessment capabilities, and much more.

We also launched a new quiz experience in Gemini, which our Research team helped to design and optimize for learning. Students (ages 18+) can ask Gemini to create custom quizzes to help them study any topic, based on their class notes or course documents, and it will offer feedback and explanations about right and wrong answers.

Explore our LearnLM prompt guide to maximize the pedagogical value of Gemini, for example, by asking it to act as a biology teacher or to adjust the difficulty level of text for a particular school grade.

As well as infusing pedagogy into Google products, we’re working with partners to bring the powerful capabilities of our LearnLM models to educational settings. Along with Kayma, we piloted the automatic assessment of both short and long-form content with thousands of students and educators in high schools in Ghana, and we’re working to scale to more students and countries.

As part of Google’s mission to make the world’s information universally accessible, we are advancing research into multilinguality to ensure that LLMs produce reliable outputs in different languages and are truly useful for everyone around the world. Two months ago, Google introduced Gemma3, and our research helped Gemma expand to over 140 languages, making it today’s best multilingual open model. At I/O, we announced that these capabilities are now available with the latest addition to the Gemmaverse, Gemma3n, a model that can run on as little as two gigabytes of RAM and is built for on-device applications. Our efforts around efficiency enable the Gemma3n model to reduce latency and be more energy consumption friendly.

To help developers build and improve multilingual models, Google Research recently introduced ECLeKTic, a novel benchmark for evaluating cross-lingual knowledge transfer in LLMs.

As LLMs grow larger and demand increases, our ability to improve model efficiency while maintaining and even elevating their quality determines our success in democratizing access to these high-performing models. Google Research has made breakthroughs in efficiency that have become industry standards, for example, our work on speculative decoding and cascades.

We have published research on factual consistency techniques and evaluations, and set the bar on factuality and grounding with features like double-check and the FACTS Grounding leaderboard, released in collaboration with Google DeepMind and Kaggle. Now, we have contributed our research to AI Mode, to meaningfully improve the experience for users.

Announced at I/O, AI Mode is Google’s most powerful AI search yet with advanced reasoning capabilities. It is rolling out to all users in the U.S., allowing people to conduct deeper research with follow-up questions and links to relevant sites. Our work on efficiency enables the models to run more reliably and serve quicker outputs, and our factuality research has improved the way AI Mode searches the web, helping to ensure that the answers provided are highly accurate and grounded in multiple sources with relevant links.

As multimodal content becomes ubiquitous, our factuality team is advancing research around multimodal factuality to ensure high accuracy standards across Google products. We improved the quality of Imagen4 on the Gemini app, the latest image model announced at I/O, which can deliver visuals with lifelike detail. For AI avatars in Vids, a new feature that enables users to create video content with their chosen AI avatars in a matter of seconds, we helped to evaluate the quality of the model and image captions. We also delivered significant enhancements to the video understanding capabilities of Gemini 2.5 models, specifically targeting high motion understanding so that Gemini is more capable of assessing human motion across health and fitness domains.

Our teams helped support the launch of the new Labs experiment, Sparkify. Bringing together the power of Gemini, MusicLM, AudioLM, and Veo, Sparkify allows users to turn any question or idea into a short and engaging animated video in the design style of their choice. The project builds on the underlying models and their factuality. Sign up for the waitlist for a chance to try it out.

As part of our long-established efforts to help reduce the devastating impacts of wildfires, Google Research has partnered with the Earth Fire Alliance, the Moore Foundation and Muon Space to develop FireSat. FireSat is a constellation of satellites built for earlier and more accurate global wildfire detection. It uses high-res multispectral satellite imagery and AI to provide near real-time insights for first responders, and to allow scientists and ML experts to study fire propagation. In March, we launched the first of over 50 satellites in the constellation. This work expands upon our wildfire boundary tracking, which makes critical information available in Search and Maps, and the synthetic Firebench dataset, which we released on the Google Cloud Platform to advance scientific research in the field.

On the Dialogues Stage, Sinead Bovell, founder of WAYE, and Julian Kelly, Senior Director, from our Quantum Hardware team, discussed the promise of quantum computing and the engineering and scientific challenges that still need to be overcome. Julian highlighted the recent advances from Google Research's Quantum AI team, including our Willow chip and progress in areas like quantum error correction. Computations that are beyond reach for classical computers can be completed on a quantum chip in minutes, paving the way for various real-world applications in the future. The potential to revolutionize fields like drug discovery and energy efficiency is becoming increasingly tangible.

We also created an interactive Quantum AI game experience for attendees on the ground at I/O: the Quantum Maze Runner. Players had to race against the clock to complete the maze, and then see how the quantum computer would solve it.

Our AI co-scientist, mentioned at I/O and developed in collaboration with Google DeepMind, is a multi-agent system based on Gemini that can synthesize information and perform complex reasoning tasks. It is designed as a collaborative tool for scientists, to aid them in creating novel hypotheses and research proposals, and to help accelerate biomedical discoveries. It has demonstrated potential in areas such as drug repurposing for acute myeloid leukemia and proposing hypotheses for novel treatment targets for liver fibrosis.

It is one of our many efforts to accelerate scientific research across the wider ecosystem. Our new Geospatial Reasoning initiative aims to advance public health, urban planning, integrated business planning, climate science and more. We’re also advancing neuroscience, with our recent publication on LICONN, the first-ever method for using commonly available light microscopes to comprehensively map neurons and their connections in brain tissue, and with the release of the Zebrafish Activity Prediction Benchmark (ZAPBench), which allows researchers to investigate the relationship between the structural wiring and dynamic neural activity across an entire vertebrate brain for the first time. We’re also advancing research into genomics to help diagnose rare diseases; REGLE is an unsupervised deep learning model that helps researchers discover associations with genetic variants. And we open sourced new DeepVariant models as part of a collaboration on Personalized Pangenome References, which can reduce errors by 30% when analyzing genomes of diverse ancestries.

The research highlighted here represents some of the ongoing work done by the Google Research teams who are driving breakthroughs across a variety of fields and bringing them to reality. In this golden age of research, the “magic cycle” between research and real-world application is increasingly faster and broader in scope, and I/O was a great opportunity to showcase how this leads to greater impact on people, businesses, science and society.

With thanks to the many teams and collaborators who have contributed to this blog and the work represented here.