Our new Perch model helps conservationists analyze audio faster to protect endangered species, from Hawaiian honeycreepers to coral reefs.One of the ways scientists protect the health of our planet’s wild ecosystems is by using microphones (or underwater hydrophones) to collect vast amounts of audio dense with vocalizations from birds, frogs, insects, whales, fish and more. These recordings can tell us a lot about the animals present in a given area, along with other clues about the health of that ecosystem. Making sense of so much data, however, remains a massive undertaking.Today, we are releasing an update to Perch, our AI model designed to help conservationists analyze bioacoustic data. This new model has better state-of-the-art off-the-shelf bird species predictions than the previous model. It can better adapt to new environments, particularly underwater ones like coral reefs. It’s trained on a wider range of animals, including mammals, amphibians and anthropogenic noise — nearly twice as much data in all, from public sources like Xeno-Canto and iNaturalist. It can disentangle complex acoustic scenes over thousands or even millions of hours of audio data. And it’s versatile, able to help answer many different kinds of questions, from “how many babies are being born” to “how many individual animals are present in a given area.”In order to help scientists protect our planet’s ecosystems, we’re releasing this new version of Perch as an open model and making it available on Kaggle.

Our novel Deep Loop Shaping method improves control of gravitational wave observatories, helping astronomers better understand the dynamics and formation of the universe.To help astronomers study the universe’s most powerful processes, our teams have been using AI to stabilize one of the most sensitive observation instruments ever built.In a paper published today in Science, we introduce Deep Loop Shaping, a novel AI method that will unlock next-generation gravitational-wave science. Deep Loop Shaping reduces noise and improves control in an observatory’s feedback system, helping stabilize components used for measuring gravitational waves — the tiny ripples in the fabric of space and time.These waves are generated by events like neutron star collisions and black hole mergers. Our method will help astronomers gather data critical to understanding the dynamics and formation of the universe, and better test fundamental theories of physics and cosmology.We developed Deep Loop Shaping in collaboration with LIGO (Laser Interferometer Gravitational-Wave Observatory) operated by Caltech, and GSSI (Gran Sasso Science Institute), and proved our method at the observatory in Livingston, Louisiana.LIGO measures the properties and origins of gravitational waves with incredible accuracy. But the slightest vibration can disrupt its measurements, even from waves crashing 100 miles away on the Gulf coast. To function, LIGO relies on thousands of control systems keeping every part in near-perfect alignment, and adapts to environmental disturbances with continuous feedback.Deep Loop Shaping reduces the noise level in the most unstable and difficult feedback loop at LIGO by 30 to 100 times, improving the stability of its highly-sensitive interferometer mirrors. Applying our method to all of LIGO’s mirror control loops could help astronomers detect and gather data about hundreds of more events per year, in far greater detail.In the future, Deep Loop Shaping could also be applied to many other engineering problems involving vibration suppression, noise cancellation and highly dynamic or unstable systems important in aerospace, robotics, and structural engineering.

We’re partnering with Commonwealth Fusion Systems (CFS) to bring clean, safe, limitless fusion energy closer to reality.Fusion, the process that powers the sun, promises clean, abundant energy without long-lived radioactive waste. Making it work here on Earth means keeping an ionized gas, known as plasma, stable at temperatures over 100 million degrees Celsius — all within a fusion energy machine’s limits. This is a highly complex physics problem that we’re working to solve with artificial intelligence (AI).Today, we’re announcing our research partnership with Commonwealth Fusion Systems (CFS), a global leader in fusion energy. CFS is pioneering a faster path to clean, safe and effectively limitless fusion energy with its compact, powerful tokamak machine called SPARC.SPARC leverages powerful high-temperature superconducting magnets and aims to be the first magnetic fusion machine in history to generate net fusion energy — more power from fusion than it takes to sustain it. That landmark achievement is known as crossing “breakeven,” and a critical milestone on the path to viable fusion energy.This partnership builds on our groundbreaking work using AI to successfully control a plasma. With academic partners at the Swiss Plasma Center at EPFL (École Polytechnique Fédérale de Lausanne), we showed that deep reinforcement learning can control the magnets of a tokamak to stabilize complex plasma shapes. To cover a wider range of physics, we developed TORAX, a fast and differentiable plasma simulator written in JAX.Now, we’re bringing that work to CFS to accelerate the timeline to deliver fusion energy to the grid. We’ve been collaborating on three key areas so far:Producing a fast, accurate, differentiable simulation of a fusion plasma.Finding the most efficient and robust path to maximizing fusion energy.Using reinforcement learning to discover novel real-time control strategies.The combination of our AI expertise with CFS’s cutting-edge hardware makes this the ideal partnership to advance foundational discoveries in fusion energy for the benefit of the worldwide research community, and ultimately, the whole world.Simulating fusion plasmaTo optimize the performance of a tokamak, we need to simulate how heat, electric current and matter flow through the core of a plasma and interact with the systems around it. Last year, we released TORAX, an open-source plasma simulator built for optimization and control, expanding the scope of physics questions we could address beyond magnetic simulation. TORAX is built in JAX, so it can run easily on both CPUs and GPUs and can smoothly integrate AI-powered models, including our own, to achieve even better performance.TORAX will help CFS teams test and refine their operating plans by running millions of virtual experiments before SPARC is even turned on. It also gives them flexibility to quickly adapt their plans once the first data arrives.This software has become a linchpin in CFS’s daily workflows, helping them understand how the plasma will behave under different conditions, saving precious time and resources.

In the rapidly evolving landscape of large language models (LLMs), the spotlight has largely focused on the decoder-only architecture. While these models have shown impressive capabilities across a wide range of generation tasks, the classic encoder-decoder architecture, such as T5 (The Text-to-Text Transfer Transformer), remains a popular choice for many real-world applications. Encoder-decoder models often excel at summarization, translation, QA, and more due to their high inference efficiency, design flexibility, and richer encoder representation for understanding input. Nevertheless, the powerful encoder-decoder architecture has received little relative attention.Today, we revisit this architecture and introduce T5Gemma, a new collection of encoder-decoder LLMs developed by converting pretrained decoder-only models into the encoder-decoder architecture through a technique called adaptation. T5Gemma is based on the Gemma 2 framework, including adapted Gemma 2 2B and 9B models as well as a set of newly trained T5-sized models (Small, Base, Large and XL). We are excited to release pretrained and instruction-tuned T5Gemma models to the community to unlock new opportunities for research and development.From decoder-only to encoder-decoderIn T5Gemma, we ask the following question: can we build top-tier encoder-decoder models based on pretrained decoder-only models? We answer this question by exploring a technique called model adaptation. The core idea is to initialize the parameters of an encoder-decoder model using the weights of an already pretrained decoder-only model, and then further adapt them via UL2 or PrefixLM-based pre-training.

An overview of our approach, showing how we initialize a new encoder-decoder model using the parameters from a pretrained, decoder-only model.

This adaptation method is highly flexible, allowing for creative combinations of model sizes. For instance, we can pair a large encoder with a small decoder (e.g., a 9B encoder with a 2B decoder) to create an “unbalanced” model. This allows us to fine-tune the quality-efficiency trade-off for specific tasks, such as summarization, where a deep understanding of the input is more critical than the complexity of the generated output.Towards better quality-efficiency trade-offHow does T5Gemma perform?In our experiments, T5Gemma models achieve comparable or better performance than their decoder-only Gemma counterparts, nearly dominating the quality-inference efficiency pareto frontier across several benchmarks, such as SuperGLUE which measures the quality of the learned representation.

Encoder-decoder models consistently offer better performance for a given level of inference compute, leading the quality-efficiency frontier across a range of benchmarks.

This performance advantage isn’t just theoretical; it translates to real-world quality and speed too. When measuring the actual latency for GSM8K (math reasoning), T5Gemma provided a clear win. For example, T5Gemma 9B-9B achieves higher accuracy than Gemma 2 9B but with a similar latency. Even more impressively, T5Gemma 9B-2B delivers a significant accuracy boost over the 2B-2B model, yet its latency is nearly identical to the much smaller Gemma 2 2B model. Ultimately, these experiments showcase that encoder-decoder adaptation offers a flexible, powerful way to balance across quality and inference speed.Unlocking foundational and fine-tuned capabilitiesCould encoder-decoder LLMs have similar capabilities to decoder-only models?Yes, T5Gemma shows promising capabilities both before and after instruction tuning.After pre-training, T5Gemma achieves impressive gains on complex tasks that require reasoning. For instance, T5Gemma 9B-9B scores over 9 points higher on GSM8K (math reasoning) and 4 points higher on DROP (reading comprehension) than the original Gemma 2 9B model. This pattern demonstrates that the encoder-decoder architecture, when initialized via adaptation, has the potential to create a more capable, performant foundational model.

Detailed results for pretrained models, illustrating how adapted models have significant gains on several reasoning-intensive benchmarks compared to decoder-only Gemma 2.

These foundational improvements from pre-training set the stage for even more dramatic gains after instruction tuning. For example, comparing Gemma 2 IT to T5Gemma IT, the performance gap widens significantly across the board. T5Gemma 2B-2B IT sees its MMLU score jump by nearly 12 points over the Gemma 2 2B, and its GSM8K score increases from 58.0% to 70.7%. The adapted architecture not only potentially provides a better starting point but also responds more effectively to instruction-tuning, ultimately leading to a substantially more capable and helpful final model.

Detailed results for fine-tuned + RLHFed models, illustrating the capabilities of post-training to significantly amplify the performance advantages of the encoder-decoder architecture.

Explore our models: Releasing T5Gemma checkpointsWe’re very excited to present this new method of building powerful, general purpose encoder-decoder models by adapting from pretrained decoder-only LLMs like Gemma 2. To help accelerate further research and allow the community to build on this work, we are excited to release a suite of our T5Gemma checkpoints.The release includes:Multiple Sizes: Checkpoints for T5-sized models (Small, Base, Large, and XL), the Gemma 2-based models (2B and 9B), as well as an additional model in between T5 Large and T5 XL.Multiple Variants: Pretrained and instruction-tuned models.Flexible Configurations: A powerful and efficient unbalanced 9B-2B checkpoint to explore the trade-offs between encoder and decoder size.Different Training Objectives: Models trained with either PrefixLM or UL2 objectives to provide either state-of-the-art generative performance or representation quality.We hope these checkpoints will provide a valuable resource for investigating model architecture, efficiency, and performance.Getting started with T5GemmaWe can’t wait to see what you build with T5Gemma. Please see the following links for more information:Learn about the research behind this project by reading the paper.Explore the models capabilities or fine-tune them for your own use cases with the Colab notebook.

AcknowledgementsWe would like to thank Zoubin Ghahramani, Helen King, Rahul Sukthankar, Raia Hadsell, and Chandu Thota for their leadership and support.This work was done thanks to the contributions of Mevan Babakar, Hannah Forbes-Pollard, Nikki Hariri, Thomas Leung, Nick Dufour, Ben Usman, Min Ma, Steve Pucci, Spudde Childs, Kate Harrison, Alanna Slocum, Reza Aghajani, Sri Rajendran, Alexey Vorobyov, Ashley Eden, Rishub Jain, Stephanie Chan, Sophie Bridgers, Michiel Bakker, Sures Kumar Thoddu Srinivasan, Tesh Goyal, and Ashish Chaudhary.We would also like to thank Kent Walker, Camino Rojo, Clement Wolf, J.D. Velazquez, Tom Lue, Ndidi Elue, Rachel Stigler, M.H. Tessler, Ricardo Prada, William Isaac, Tom Stepleton, Zoe Darme, Gail Kent, Vincent Ryan, Aaron Donsbach, Abhishek Bapna, Verena Rieser, Christian Plagemann, Anca Dragan, Joelle Barral, Edward Grefenstette, Sara Mahdavi, Sven Gowal, Florian Stimberg, Christopher Savcak, Allison Garcia, Eve Novakovic, Armin Senoner, Arielle Bier, and the greater Google DeepMind and Google teams for their support, help, and feedback.

AcknowledgementsWe thank the International Mathematical Olympiad organization for their support.This project was a large-scale collaboration, and its success is due to the combined efforts of many individuals and teams. Thang Luong led the overall technical direction for IMO 2025 effort and co-led with Edward Lockhart on the overall coordination.The leads and key contributors of the IMO 2025 team are the following; Dawsen Hwang, Junehyuk Jung, Jonathan Lee, Nate Kushman, Pol Moreno, Yi Tay, Lei Yu, Golnaz Ghiasi, Garrett Bingham, Lalit Jain, Vincent Cohen-Addad and Theophane Weber, Ankesh Anand, Steven Zheng, Vinh Tran, Vinay Ramasesh, Andreas Kirsch, Jieming Mao, Zicheng Xu, Wilfried Bounsi, Vahab Mirrokni, Hoang Nguyen, Fred Zhang, Mahan Malihi, Yangsibo Huang, Yuri Chervonyi, Trieu Trinh, Junsu Kim, Mirek Olšák, Marcelo Menegali, Xiaomeng Yang, Richard Song, Miklós Z. Horváth, Aja Huang, Goran Žužić.The advanced Gemini model with Deep Think for IMO was built on foundational research from the Deep Think team with sponsorship of the GDM Thinking area, and corresponding post-training efforts including; Archit Sharma, Shubha Raghvendra, Tong He, Pei Sun, Tianhe (Kevin) Yu, Eric Ni, Siamak Shakeri, Hanzhao (Maggie) Lin, Cosmo Du, Sid Lall, Le Hou, Yuan Zhang, Yujing Zhang, Yong Cheng, Luheng He, and Chenxi Liu.This effort was advised by Quoc Le and Pushmeet Kohli, with program management from Kristen Chiafullo and Alex Goldin.We’d also like to thank our experts for providing data and evaluations: Insuk Seo (lead), Jiwon Kang, Donghyun Kim, Junsu Kim, Jimin Kim, Seongbin Jeon, Yoonho Na, Seunghwan Lee, Jihoo Lee, Younghun Jo, Yongsuk Hur, Seongjae Park, Kyuhyeon Choi, Minkyu Choi, Su-Hyeok Moon, Seojin Kim, Yueun Lee, Taehun Kim, Jeeho Ryu, Seungwoo Lee, Dain Kim, Sanha Lee, Hyunwoo Choi, Aiden Jung, Youngbeom Jin, Jeonghyun Ahn, Junhwi Bae, Gyumin Kim, Nam Dung Tran, Quoc Ba Can Vo, Van Huyen Nguyen, Tuan Anh Nguyen, Thanh Dat Vo, Nguyen Nam Hung Tran, Van Khai Luong, Son Vu, Son Tra Dao, Dai Dinh Phong Tran, Thanh Dat Le, Cheng-Chiang Tsai, Kari Ragnarsson, Kiat Chuan Tan, Yahya Tabesh, Hamed Mahdavi, Azin Nazari, Chu-Lan Kao, Steven Creech, Tony Feng, Daogao Liu, and Ciprian Manolescu.Further thanks to the following people for support, collaboration, and advice; Omer Levy, Timothy Lillicrap, Jack Rae, Yifeng Lu, Heng-tze Cheng, Denny Zhou, Ed Chi, Vahab Mirrokni, Tulsee Doshi, Madhavi Sewak, Melvin Johnson, Fernando Pereira, Benoit Schillings, Koray Kavukcuoglu, Oriol Vinyals, Jeff Dean, Demis Hassabis, Sergey Brin, Jessica Lo, Sajjad Zafar, Tom Simpson, Jane Labanowski, Andy Forbes, Sean Nakamoto, Jonathan Lai, Fabian Pedregosa, Samuel Albanie, Alex Zhai, Sara Javanmardi, Divy Thakkar, YaGuang Li, Nigamaa Nayakanti, Chenjie Gu, Chenkai Kuang, Swaroop Mishra, Filipe Miguel de Almeida, Silvio Lattanzi, Ashkan Norouzi Fard, Tal Schuster, Ziwei Ji, Honglu Fan, Xuezhi Wang, Aditi Mavalankar, Tom Schaul, Rosemary Ke, Xiangzhuo Ding, Adam Brown, Emanuel Taropa, Charlie Chen, Joe Stanton, Cip Baetu, Alvin Abdagic, Federico Lebron, Ioana Mihailescu, Soheil Hassas Yeganeh, Ashish Shenoy, and Minh GiangFinally, we thank Prof Gregor Dolinar from the IMO Board for the support and endorsement.The IMO have confirmed that our submitted answers are complete and correct solutions. It is important to note that their review does not extend to validating our system, processes, or underlying model (see more).