A new framework for scientific artificial intelligence, dubbed Intern-S2-Preview, has been proposed by researchers. This series of scientific agentic foundation models is designed to enhance multimodal scientific understanding, reasoning, and generation, with a particular focus on long-horizon tasks. The development addresses the growing need for AI systems capable of processing heterogeneous scientific evidence, interacting with scientific tools, and sustaining progress over extended research periods.
The training pipeline for Intern-S2-Preview begins with scientific multimodal pre-training. This initial phase involves processing rendered scientific documents, interleaved image-text data, and diverse scientific corpora. Following this pre-training, a unified post-training pipeline is applied. This pipeline includes supervised fine-tuning, scalable multi-task reinforcement learning (RL), and both black- and white-box agentic RL, as well as on-policy distillation.
Practical techniques are integrated to improve the stability and efficiency of the training process. These include 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 architectural level, Intern-S2-Preview-397B, a 397 billion parameter model, extends time series modeling capabilities for efficient long-sequence understanding and numerical forecasting. Additionally, a separate component called the Memory Decoder is studied as a memory-augmented path for rapid scientific specialization. This modular approach allows for specialization without altering the core frozen 397B backbone of the model.
Evaluations conducted across scientific, multimodal, agentic, and general-purpose benchmarks indicate that Intern-S2-Preview-397B achieves competitive or leading results in several settings. For instance, the time series modules have shown improvements in scientific signal understanding and forecasting on the SciTS benchmark. The Intern-MemDec-4B extension, a specific application of the Memory Decoder, has improved the average score on Biology-Instructions from 56.92 to 60.32 without modifying the underlying 397B model.
The development of Intern-S2-Preview represents a move towards AI systems that can actively participate in the scientific process. By combining visual perception, numerical forecasting, and tool-based interaction, the model offers a foundation for more autonomous scientific workflows. The modularity provided by the Memory Decoder is intended to help the model keep pace with the evolving scientific landscape, allowing researchers to train lightweight memory modules for new fields rather than re-training massive foundation models.
This work builds upon previous research, such as Intern-S1-Pro, a trillion-scale scientific multimodal foundation model. Intern-S2-Preview is presented as an evolution, incorporating upgrades to its time series modules and post-training pipeline.
The framework is designed to support complex scientific workflows, including autonomous experimental design and software engineering for scientific applications.