This @GoogleDeepMind's paper is really worth your time It's on how to help Transformers not lose the right context on the way to the final answer. For this, the researchers introduce Recirculation: Normally, information passes through Transformer layers once. Recirculation changes that flow: → Some of what the model figures out in deeper layers is passed back to earlier layers and used when processing the next input. For example, once the model understands that “bank” means river bank in a fishing context, that interpretation can stay available when it later gets a question about an ATM. So here is how recirculation works: 1. The model processes the input normally. 2. Deeper layers build a more contextualized representation. 3. A small part of that activation is mixed back into a shallower layer. 4. The next input is processed with this updated state. The weights stay frozen, and you're changing how information flows through the model at inference time, not retraining it. This method really works in practice: - recirculation reduced contextualization errors by 60% - reduced perplexity by 23% - improved GSM8K accuracy by 21% - improved performance on several other tasks And now, the most interesting question: is this an alternative to Chain-of-Thoughts? Not really. CoT adds computation through generated reasoning tokens; recirculation helps the model keep track of what it has already understood internally. It’s also different from looped Transformers: they repeat the same layers, effectively adding depth, while recirculation feeds deeper representations back into earlier processing. So instead of asking the model to reason out loud, let it reuse more of what it has already figured out internally.

Join @GoogleDeepMind Principal Engineer @__apf__ to walk through how Gemini Spark helps simplify your daily workflows. Powered by Gemini 3.5 Flash, Spark builds upon Gemini's ability to connect with @GoogleWorkspace apps like Docs and Gmail to execute complex tasks.

Gemini 4 Argon by @GoogleDeepMind just went head-to-head with top frontier models across a set of generations. At 33% lower cost per task than GPT-6.1 Sol, and 70% lower than Claude Opus 5.5, see first impressions from @petergostev on how Gemini 4 Argon compares.

our amazing research team @GoogleDeepMind just launched the DeepMind Institute (DMI), a platform to spur the interdisciplinary research, collaboration and debate required to answer the AGI era's most critical technical & societal questions… tackling all the questions like - what will we value, and how will AGI impact what it means to be human? how do we safely build and govern AGI systems and the communities of agents they will form? which institutions and policies will society need to adapt to AGI or reimagine altogether?

New from @GoogleResearch and @GoogleDeepMind: AMIE can now conduct real-time video consultations. In a randomized study using simulated consultations, it met clinical performance benchmarks, showing AI’s potential to expand telehealth access. Learn more:

Congratulations to the @GoogleDeepMind authors of "Asynchronous Methods for Deep Reinforcement Learning", recipient of the #ICML2026# Test of Time Award. This work shows that asynchronous actor-critic succeeds on a wide variety of continuous motor control problems as well as on a new task of navigating random 3D mazes using a visual input.

New from @GoogleDeepMind: Announcing AI co-clinician, a research initiative exploring the potential for real-time multimodal AI as an assistive component of the care team.

Had early access to Genie 4 from @GoogleDeepMind. This is a fully generated, interactive world from a single starting prompt (environment + character) 🤯 Feels like a massive step forward - you can now prompt to change the world and interact with it. With native audio!