Can computational power accelerate scientific discovery itself?

ASI-ARCH is the first demonstration of fully autonomous neural architecture discovery at scale. A three-module agent system (Researcher → Engineer → Analyst) with persistent memory conducted 1,773 autonomous experiments over 20,000 GPU hours, discovering 106 state-of-the-art linear attention architectures.

The most significant finding is not the architectures themselves but the discovery of the first empirical scaling law for scientific discovery: architectural breakthroughs scale computationally. This transforms research progress from a human-limited process to a computation-scalable one.

Key mechanisms enabling this:

• Cognition base: ~100 seminal papers structured into scenario/algorithm/historical-context entries. The Analyst's problem summaries query this base by semantic similarity, ensuring relevant prior knowledge surfaces at the right time.
• Self-revision: Unlike previous NAS approaches that discard failed runs, ASI-ARCH requires agents to fix their own implementation errors. Iterative debugging continues until training succeeds — preserving promising ideas that would be lost to simple coding mistakes.
• Quality assurance: Real-time monitoring of training logs detects anomalies (excessive training duration, abnormally low loss) and terminates runs proactively, reporting issues for revision.
• Two-tier evolution: Parents from top-10 performers serve as bases for modification; references from positions 11-50 provide diversity. This balances exploitation with exploration.
• Novelty check: Embedding-based similarity search against historical motivations prevents redundant proposals.

The "AlphaGo Move 37" analogy is deliberate: AI-discovered architectures demonstrate emergent design principles that systematically surpass human baselines and illuminate previously unknown pathways — design insights invisible to human designers.

The exploration-then-verification two-stage strategy is practical: broad exploration on small models (cheap) narrows to rigorous validation on larger models (expensive). This is the same ladder-of-scales principle that makes research affordable.

ADAS as precursor (from Arxiv/Agents): The Automated Design of Agentic Systems (ADAS) research formulated the explicit thesis that "the history of machine learning teaches us that hand-designed solutions are eventually replaced by learned solutions." ADAS demonstrated that agents can be defined in code and new agents automatically discovered by a meta-agent programming ever better ones. While ASI-ARCH focuses on neural architecture search, ADAS extends the principle to agentic system design itself — the meta-agent discovers novel building blocks and combinations for agent architectures. This confirms the scaling law operates at multiple levels: not just architectures, but the design of the systems that discover architectures.

Source: Novel Architectures, Agents