Reproduce “ATHENA: Agentic Team for Hierarchical Evolutionary Numerical Algorithms” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Juan Diego Toscano, Daniel T. Chen, George Em Karniadakis

ATHENA, an agentic framework, emulates the scientific research process through a knowledge-driven contextual bandit approach for numerical algorithms. The system autonomously derives analytical solutions, orchestrates high-order numerical solvers, and achieves high accuracy on PIML benchmarks, often surpassing expert-authored solutions while demonstrating efficient, sublinear reward improvement.

Key insights

  • Modeling scientific discovery as a contextual bandit problem provides a principled mechanism for coordinating agents and managing information flow to achieve efficient, guided search.
  • The clear separation of conceptual policy (high-level strategy) from numerical realization (implementation) allows for robust diagnosis, reformulation, and repair of computational strategies using expert-derived blueprints.
  • Interpreting multimodal diagnostic observations (plots, logs) in conjunction with fundamental scientific principles enables the system to identify failure modes, assess physical plausibility, and propose coherent, corrective adjustments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.