Scientific discovery increasingly requires AI systems that can reason over scientific evidence of heterogeneous modalities, interact with scientific tools and environments, and sustain progress across long task horizons. We present Intern-S2-Preview, a series of scientific agentic foundation models designed to support multimodal scientific understanding, reasoning, generation, and long-horizon tasks. The training pipeline begins with scientific multimodal pre-training over rendered scientific documents, interleaved image-text data, and diverse scientific corpora. Starting from the pretrained checkpoint, we apply a unified post-training pipeline consisting of supervised fine-tuning, scalable multi-task reinforcement learning (RL), black- and white-box agentic RL, and on-policy distillation. This pipeline is supported by practical techniques that improve rollout and training stability and efficiency, including partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly for agentic tasks. At the architecture level, Intern-S2-Preview-397B extends time series modelling from efficient long-sequence understanding to numerical forecasting, while Memory Decoder is studied as a separate memory-augmented path for rapid scientific specialization without modifying the frozen 397B backbone. Evaluations across scientific, multimodal, agentic, and general-purpose benchmarks show that Intern-S2-Preview-397B achieves competitive or leading results in multiple settings. The time series modules improve scientific signal understanding and forecasting on SciTS, while the separate Intern-MemDec-4B extension improves the Biology-Instructions average score from 56.92 to 60.32 without modifying the frozen 397B backbone.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at Intern-S2-Preview, a system designed to help AI agents tackle complex scientific projects.
Sam: This research addresses a core challenge in AI: how do you build a system that can handle multi-step scientific work—reading papers, using software tools, running simulations—without losing its general reasoning skills along the way? The authors' answer is to keep the main model completely frozen and attach a separate, lightweight module alongside it.
Alex: So instead of retraining the entire model every time it needs to learn a new field, they just plug in a kind of specialized reference guide?
Sam: Exactly. Think of it like a scientist who spent years building up their core education. You wouldn't want to erase that foundation every time they started a new project. Instead, you hand them a detailed manual specific to that project. The main model—which the paper calls the "backbone"—keeps all its general reasoning intact, while a separate component called the Memory Decoder provides the domain-specific knowledge it needs.
Alex: And why is keeping the backbone frozen so important?
Sam: There's a well-documented problem in AI called catastrophic forgetting. When you update a model's internal settings to teach it something new, it tends to overwrite what it already knew. It's a bit like cramming for one exam and blanking on everything from the previous term. By keeping the backbone frozen and routing new knowledge through a separate module, the system avoids that problem entirely.
Alex: So the core reasoning stays intact, and the specialized knowledge lives somewhere else. How does the system know which one to lean on at any given moment?
Sam: That's where an intelligent component called a "router" comes in. It acts as a kind of traffic controller, deciding in real time how much weight to give the main model versus the memory module, depending on the question being asked. This is part of a broader shift the paper describes toward what they call "agentic" systems—AI that doesn't just answer a single question but actively plans and executes tasks over long stretches of time.
Alex: That's a meaningful distinction. A single question is one thing, but a real scientific workflow might involve dozens of steps, each depending on the last.
Sam: Right. And that's where things get technically interesting. Because these long tasks are expensive to run, the researchers use a technique called speculative decoding. Imagine a fast but less experienced assistant who drafts the next few steps of a solution. The main model then checks that draft. If it's correct, great—you've saved a lot of time. If it's wrong, the draft gets discarded and the main model takes over for that step.
Alex: So the assistant is essentially doing the rough work, and the expert is doing quality control?
Sam: That's a good way to put it. They call the checking process a rejection-sampling procedure. Nothing incorrect makes it into the final output, so accuracy is preserved. The speed gain comes from the fact that the main model only has to verify, not generate from scratch every single time.
Alex: What about the visual side of science? A huge amount of scientific knowledge lives in charts, diagrams, and figures—not just text.
Sam: That's a real problem the paper takes seriously. A lot of scientific knowledge is effectively locked inside figures that text-only systems simply can't access. To address this, the model is trained on rendered scientific documents—actual PDFs with charts and equations—so it learns to understand visual layout, not just words.
Alex: But not every figure in a paper is actually useful. Some are just decorative.
Sam: Exactly, and that's why they build in what they call visual-gain filtering. The system checks whether a given diagram actually helps clarify the surrounding text. If the image adds meaningful information—what they call "gain"—the model uses it. If it's just decorative, the system ignores it. It's a way of filtering signal from noise before the model even starts reasoning.
Alex: That sounds like a genuinely practical design choice. It keeps the agent focused on what actually matters in a document.
Sam: And it connects to the broader goal of the system. By combining visual understanding, tool use, and specialized memory, the paper argues this approach can close the gap between static knowledge—what a model learned during training—and the kind of iterative, tool-heavy workflows that define real laboratory research. Whether it fully achieves that is something the field will continue to evaluate, but the architecture is a considered attempt at the problem.
Alex: It's a thoughtful set of design choices. Rather than one dramatic solution, it's a collection of careful decisions—each one addressing a specific weakness in how AI systems currently handle scientific work.
Sam: That's a fair characterization. The system is designed to be modular and adaptable, which means individual components can be improved or swapped out as the field develops. It's less a finished product and more a framework for thinking about what a capable scientific AI agent actually needs.
Alex: Thanks for walking us through it. And thanks to everyone listening to ResearchPod.