| Contents | Intelligence is commonly understood as the ability to acquire knowledge, adapt to unfamiliar situations, and solve new problems. Large language models exhibit this capacity by inferring task-relevant knowledge from textual context and applying it to new tasks. Yet intelligence need not be confined to language. Scientific and social systems often reveal themselves numerically before we can fully describe them in words. We call the ability to acquire and apply knowledge from numerical context “numerical intelligence” and view it as a foundational pillar of artificial intelligence alongside linguistic intelligence.
We proposed “In-Context Operator Learning” paradigm and the corresponding model “In-Context Operator Networks” (ICON) for numerical intelligence. We will show how a single ICON model (without fine-tuning) manages multiple families of PDE problems. Apart from the model itself, we can also design “harness” for ICON due to the flexibility of the context. We will introduce building blocks, including “chain of operators” .
To accelerate our research, we built Evolving Ensemble of Agents (EvE), a decentralized agent system that co-evolves with downstream task solutions. We will show how EvE helps ourselves design ICON architectures and harness systems, and beyond.
|
