Intern-S2-Preview: Scientific Agentic Foundation Model
The paper introduces Intern-S2-Preview, a series of scientific agentic foundation models trained through a specialized multimodal pre-training and unified post-training pipeline featuring advanced reinforcement learning and memory-augmented architectures to achieve state-of-the-art performance in scientific reasoning, tool interaction, and long-horizon tasks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a robot to be a scientist. In the past, we gave robots big libraries of text and asked them to answer questions like a quiz show contestant. But real science isn't a quiz; it's a messy, long adventure. It involves looking at weird pictures of cells, reading graphs that wiggle like heartbeats, using computer tools to run experiments, and thinking through a problem for hours without getting lost. This paper is about building a new kind of "super-brain" for robots that doesn't just memorize facts, but actually does science. It's designed to understand complex images, predict future numbers, and act like a detective that can use tools to solve mysteries over a long period of time. The big idea is that to be a true scientific assistant, an AI needs to be able to juggle different types of information (like text, images, and numbers) and keep working steadily on a hard problem without giving up or getting confused.
The team behind this project, from the Shanghai AI Laboratory, introduces a new family of AI models called Intern-S2-Preview. Think of this model as a massive, 397-billion-parameter "generalist" scientist who has read almost every scientific paper ever written. But instead of just sitting there knowing things, this model is built to act. It can look at a scientific diagram, understand the story behind it, and then use computer tools to run simulations or write code to test a hypothesis.
Here is how they built this super-scientist:
First, they taught the model by showing it millions of scientific documents, but not just the text. They showed it the actual pages, with all the pictures, charts, and equations in their original layout. It's like teaching a student by handing them the whole textbook, including the diagrams, rather than just reading the text aloud. This helps the model understand how scientists actually present their ideas.
Then, they gave the model a special "time-travel" skill. Most AI models are bad at predicting what happens next in a sequence of numbers (like weather patterns or stock prices). Intern-S2-Preview got a dedicated upgrade to handle these "time series" data. It can now look at a long line of numbers and accurately forecast what will happen next, which is crucial for things like predicting climate change or analyzing brain signals.
One of the coolest tricks they used is called the Memory Decoder. Imagine the main 397-billion-parameter brain is a brilliant generalist who knows a little bit about everything. But what if you need an expert in just one tiny field, like biology? Instead of retraining the whole giant brain (which would be slow and might make it forget other things), they attached a small, specialized "notebook" called the Memory Decoder. This notebook holds specific biology knowledge. When the AI needs to answer a biology question, it consults this notebook. The result? The model's biology scores jumped from 56.92 to 60.32, while its general knowledge stayed exactly the same. It's like having a genius friend who can instantly pull out a specialized manual whenever you ask a specific question, without needing to go to school for a new degree.
The team also taught the model how to be a long-horizon agent. This means the AI can plan a multi-step mission. Instead of just answering "What is the capital of France?", it can be asked to "Design a new drug." It might then: 1) search for existing research, 2) write code to simulate the drug's behavior, 3) run the simulation, 4) realize it failed, 5) adjust the code, and 6) try again. The paper shows that by using a special training method called "reinforcement learning," the model learned to handle these long, complex workflows without getting stuck or making silly mistakes.
When they tested Intern-S2-Preview-397B, it did incredibly well. On tests measuring how well it understands biology, chemistry, and physics, it beat many other top models. It was particularly good at reading scientific charts and predicting time-series data, often outperforming models that are much larger or specialized only for those tasks. For example, on a test called Biology-Instructions, it scored 56.92, and with the Memory Decoder attached, it improved to 60.32. On general reasoning tests like MMLU-Pro, it scored 89.75, which is the best among open-source models.
However, the authors are careful to call this a "Preview." It's a powerful new tool, but it's not perfect yet. They note that while the model is great at connecting reasoning to action, there is still work to be done to make it even more reliable for very long, complex scientific workflows. They also emphasize that this model is a foundation—a starting point that can be adapted with different "memory notebooks" for different scientific fields without needing to rebuild the whole brain.
In short, this paper presents a new kind of AI that is ready to move beyond just answering questions and start doing the actual work of scientific discovery, from reading complex graphs to running long experiments, all while keeping its general knowledge intact and its specialized skills sharp.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.