Thinking adds a variable-length reasoning stage so Gemini can spend more computation on difficult requests.
2
Reinforcement learning teaches the model when and how to use thinking, including decomposition, self-correction, code drafting, calculations, and tool use.
3
Gemini is moving toward deeper and more efficient inference, including high-budget parallel reasoning for difficult multimodal, code, and mathematics problems.
Summary
Jack Rae presents thinking in Gemini as a response to a limitation in current language models: they usually apply a fixed amount of computation between a user request and the final answer. Thinking adds an iterative stage before the answer, allowing the model to decide how much inference computation a problem needs. Gemini learns this behavior through reinforcement learning, receiving rewards based on whether it solves tasks correctly. Rae describes emergent behaviors such as testing hypotheses, rejecting failed approaches, breaking problems into parts, drafting code, and using tools. He argues that test-time scaling adds another way to improve models alongside pre-training and post-training. Thinking also gives developers a continuous budget between cost and capability. Future work includes reducing overthinking, increasing efficiency, and using much deeper chains of thought for hard tasks such as mathematics, coding, and research.
Progress often follows the discovery of an intelligence bottleneck
Rae frames progress in language models as a cycle of finding a limitation and developing a way around it. Claude Shannon's 1948 two-gram model was limited by its small amount of data and simple statistics. Later n-gram systems scaled to trillions of tokens, but their short-context design had exponential storage costs. Recurrent neural networks compressed the past into a state, which allowed longer context but lost information because the state had a fixed size. Attention addressed that problem by keeping past representations available and selecting among them during computation. Rae places today's thinking systems in the same pattern: current models are powerful, but their response computation is usually fixed.
Thinking gives Gemini a variable amount of computation at inference time
Rae describes test-time compute as the computation applied to a user's particular question after the request has been converted into tokens. Ordinary language models use a fixed amount of this computation before producing an answer. Making the model larger adds computation, but users may want a much wider range, from a quick response to thousands or millions of additional reasoning steps. Thinking inserts a stage before the final answer. The model can loop through this stage for thousands or tens of thousands of iterations, then decide when to commit to a response. The loop is dynamic, so simpler requests can receive less computation and harder requests can receive more.
Reinforcement learning teaches the model how to use its thinking stage
After pre-training Gemini, Google uses reinforcement learning on many tasks. The model receives positive or negative rewards based on whether it solves each task correctly, and this signal is propagated through the thinking loop. Rae says the researchers were initially unsure how much structure to impose on a reasoning stage. They then saw the model develop useful behavior from the general training recipe. In one integer-prediction example, it proposed a hypothesis, tested it, rejected the formula when it failed, and tried another approach. Rae also lists learned strategies such as decomposing problems, exploring several solutions, drafting code fragments, performing intermediate calculations, and using tools.
Test-time scaling adds to pre-training and post-training
Rae argues that thinking gives developers another way to improve model performance. Pre-training can scale data and model size, while post-training can scale the quality and diversity of human feedback. Thinking scales the amount of computation used for an individual request. He says these approaches can work together and describes Gemini's recent models as showing better reasoning performance alongside higher test-time compute. He points to math, code, and science evaluations as areas where this relationship appears across the Gemini lineage, from a version that did not launch with thinking to Gemini 2.5 Pro.
Thinking budgets let developers trade cost for capability more gradually
Before thinking, developers mainly chose among discrete model sizes to decide how much quality and cost they wanted. Thinking budgets provide a more continuous control. Developers can select a smaller or larger reasoning budget for a class of tasks, rather than switching between a few fixed model choices. Rae says thinking budgets are available in the Flash and Pro models in the 2.5 series. This lets an application use more computation when the task justifies it and gives developers a more granular cost-to-performance choice.
The next challenge is making reasoning efficient enough to use automatically
Rae says Google wants thinking to adapt without requiring developers to tune it constantly. One problem is overthinking, where a model spends more computation than a task needs. The research goal is to make the models use their thoughts efficiently while keeping costs down. At the same time, Google is investigating deeper thinking, which increases inference computation for difficult problems. He describes Deep Think as a high-budget mode built on Gemini 2.5 Pro for queries that can run asynchronously. It uses deeper and parallel chains of thought that can be combined into a stronger response.
Deep Think targets difficult mathematics, code, and multimodal tasks
Rae says Deep Think improves performance on very difficult multimodal, coding, and mathematics problems. He gives USA Math Olympiad as an example. Gemini 2.5 Pro was around the 50th percentile of participants in the cited comparison, while Deep Think reached about the 65th percentile. He explains the method through an algebra proof example: the model explores several ideas, including proof by contradiction, Rolle's theorem, and Newton's inequalities, then integrates them into a proof. Rae also expects longer inference to help with open-ended coding. He describes Gemini generating the setup, algorithm, and Atari emulator needed to recreate work based on DeepMind's original DQN paper.
Rae wants models to derive more from small amounts of source information
Rae compares the direction of thinking with Srinivasa Ramanujan's mathematical work. Ramanujan had limited contact with the wider mathematical community and worked from a small set of source material, yet spent extensive time developing theories and producing new mathematics. Rae wants models to become similarly data-efficient by thinking much longer from a limited knowledge base. He describes a future in which models use millions or more inference tokens to build knowledge and intermediate artifacts, with the aim of extending the frontier of human understanding.
"Thinking in Gemini mechanically, I'm sure almost everyone in this room is familiar with this general process where we will now have a model and we insert a thinking stage that the model can emit some additional text before it decides to emit a final answer."07:16
Who should watch
You are building an application where some requests deserve more computation than others, and you want a continuous reasoning budget rather than a choice between fixed model sizes.
You want to understand how Gemini's thinking stage is trained and what behaviors reinforcement learning can produce.
You work on difficult mathematics, coding, research, or multimodal tasks where asynchronous, high-budget inference could improve the answer.
