Source Code for Navigable semantic spaces enable self-improving scientific agents
Authors/Creators
Description
Abstract:
Scientific methods are commonly developed for classes of problems, allowing knowledge to transfer across structurally related cases. LLM-based agentic systems can execute scientific workflows, but their problem--method relationships remain implicit, leaving method selection dependent on generation and limiting the connections that can be transferred or discovered. We introduce GRAFT--ATHENA, an agentic framework whose expandable probabilistic structure encodes admissible problems and methods through semantic characteristics and dependencies. Graph factorization makes this structure tractable, while semantic fingerprints make scientific similarity measurable, allowing the framework to retain and learn from evaluated experience across problems. Across scientific families, it matched or exceeded expert baselines, attaining near-machine-precision losses, clinically consistent blood-rheology trends, and a high-order solver for flow over the Apollo Command Module. GRAFT--ATHENA also proposed a verifiable regularization strategy for ill-posed in vivo brain-flow reconstruction, developed a spectrally convergent physics-informed architecture, and derived machine-checked universal-approximation theorems for commonly used scientific machine-learning architectures. Scientific structure therefore enables transferable, extensible, and verifiable agentic discovery.
Contents:
The associated file contains a README.md file with specific instructions on how to run and reproduce the results if the paper.
Files
GRAFT-ATHENA-submission.zip
Additional details
Related works
- Is supplement to
- Publication: arXiv:2605.11117 (arXiv)