Intern-S2-Preview: Scientific Agentic Foundation Model
TLDR
Intern-S2-Preview is a scientific agentic foundation model series for multimodal understanding, reasoning, generation, and long-horizon tasks, trained via pre-training and unified post-training with RL.
Reasoning
The paper presents a comprehensive training pipeline and architecture for a scientific agentic model, with evaluations across multiple benchmarks. However, the abstract is truncated and lacks specific quantitative results, making it hard to assess actual performance; the contribution is broad rather than focused on a specific automation task.
Read-first score
Read-first score 48.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.
Field roles
Rank sensitivity
Stability: volatile; rank range: 65.
Keyword Scores
Deep Analysis
Innovations
- Scientific multimodal pre-training on rendered scientific documents, interleaved image-text data, and diverse scientific corpora
- Unified post-training pipeline combining supervised fine-tuning, scalable multi-task RL, black- and white-box agentic RL, and on-policy distillation
- Practical techniques for rollout and training stability: partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly
- Extension of time series modelling from efficient long-sequence understanding to numerical forecasting
- Memory Decoder as a separate memory-augmented path for rapid scientific specialization without modifying the frozen backbone
Methodology
Intern-S2-Preview is trained via scientific multimodal pre-training on rendered documents, interleaved image-text, and scientific corpora, followed by a unified post-training pipeline of supervised fine-tuning, multi-task RL, agentic RL, and on-policy distillation. The 397B model incorporates time series modules for forecasting, and a separate Memory Decoder path enables domain specialization without backbone modification.
Key Results
Intern-S2-Preview-397B achieves competitive or leading results on scientific, multimodal, agentic, and general-purpose benchmarks. Time series modules improve scientific signal understanding and forecasting on SciTS, and the Intern-MemDec-4B extension raises the Biology-Instructions average score from 56.92 to 60.32 without altering the frozen backbone.