Hugging Face has released a research paper introducing Intern-S2-Preview, a series of scientific agentic foundation models designed to handle multimodal scientific understanding, reasoning, generation, and long-horizon tasks. The models aim to address the growing need for AI systems that can reason over heterogeneous scientific evidence, interact with tools and environments, and sustain progress over extended task horizons.
The training pipeline begins with scientific multimodal pre-training on rendered scientific documents, interleaved image-text data, and diverse scientific corpora. From this pretrained checkpoint, the team applies a unified post-training pipeline that includes supervised fine-tuning, scalable multi-task reinforcement learning (RL), black- and white-box agentic RL, and on-policy distillation. Practical techniques such as partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly are employed to improve rollout and training stability and efficiency.
At the architecture level, Intern-S2-Preview-397B extends time series modeling from efficient long-sequence understanding to numerical forecasting. Additionally, a separate Memory Decoder path is studied 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.