Scientific Computing and Intelligence Group (Scaling Group) is in the Department of Mathematics at National University of Singapore, led by Dr. Liu Yang (杨柳).
Intelligence is commonly understood as the ability to acquire and apply knowledge, adapt to unfamiliar situations, and solve new problems. Yet intelligence need not be confined to language.
Our work centers on numerical intelligence: the ability to acquire and apply knowledge from numerical context. We develop In-Context Operator Networks (ICON) as a framework that enables frozen models to infer and apply relations from numerical context across scientific and social systems. We also explore the synergy between linguistic and numerical intelligence. To accelerate this research, we develop self-evolving agents (try our EvE) for automated discovery of algorithms and beyond.
We are looking for passionate new students, postdocs, and visiting scholars to join the team!
Jul 30 2026
Our paper A Foundation Model of Numerical Intelligence with Cross-Disciplinary Generalization is now available on arXiv. UNICON is a single frozen model that infers system-specific predictive relations from graph-based numerical examples and approaches specialist performance across disciplines absent from training.
Jun 18 2026
Our paper Agentic Symbolic Search: Characterizing PDEs Beyond Hand-crafted Expressions, Meshes, and Neural Networks is now available on arXiv.
Jun 10 2026
Our paper Harness In-Context Operator Learning with Chain of Operators is now available on arXiv.
Jun 10 2026
Our paper VICX: Generalizable Robot Manipulation via Video Generation and In-Context Operator Network (Project) is now available on arXiv.
Jun 07 2026
Our paper Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control is now available on arXiv.
May 09 2026
Our paper Evolutionary Ensemble of Agents (GitHub) is now available on arXiv. We introduce EvE: a decentralized evolutionary ensemble of coding agents co-evolving with code repositories.
Apr 28 2026
Our paper In-Context Modeling as a Retrain-Free Paradigm for Foundation Models in Computational Science is now available on arXiv.
Apr 25 2026
Our paper Escher-Loop: Mutual Evolution by Closed-Loop Self-Referential Optimization is now available on arXiv.
Mar 14 2026
We released icon-core, an open-source research infrastructure package for In-Context Operator Network (ICON) development. It provides standard models and algorithms, benchmark datasets, training pipelines, and development utilities, featuring built-in support for AI coding assistants like Claude Code.
... see all news