Alphabet upgrades Gemini 3 Deep Think for math, science research

Google’s parent company Alphabet has upgraded its Gemini Deep Think AI model to improve performance in math and science research, according to a blog post.

The specialized reasoning model, developed in collaboration with researchers, aims to assist scientists in transitioning from theoretical to practical applications.

The model leverages Google Search to reduce inaccuracies and wrongful citations during research.

Named Gemini 3 Deep Think, it can support research in chemistry, computer science, and physics.

The update is part of a broader effort by AI developers to create tools capable of handling complex coding and scientific tasks.

Google also developed an autonomous research agent called Aletheia, which can collaborate with humans and recognize when it cannot solve a problem.

The company states that some research papers have been published based on the new technology.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Aletheia’s research breakthroughs are impressive, but still an exception

The company says some papers were published with its autonomous research agent, Aletheia, yet a closer look reveals limits.
On 700 unsolved math problems, evaluators found Aletheia’s proposed solutions were wrong in 68.5% of cases 1.
Just 6.5% of responses were mathematically correct and answered the question as asked, while many reframed the prompt to make it trivial 1.
On a benchmark of PhD-level problems, the agent replied to fewer than 60% of prompts, which caps its reach on harder tasks 2.
The system can still produce novel work, including a fully autonomous paper on “eigenweights,” but those wins remain rare 2.

This technology is reshaping the scientific process itself, creating new bottlenecks

Autonomous research agents like Aletheia do more than assist scientists, they reshape day-to-day research routines.
Some teams now ask the AI for a proof and a refutation at the same time to reduce confirmation bias, a method called “balanced prompting” 1.
In some projects, the AI sketches the proof strategy, then human researchers work through the technical steps, flipping the usual split of labor 1.
Faster output could trigger a peer review crunch.
If AI drives up paper volume, human verification becomes the choke point and academic publishing takes the strain 1.