DASyR-LLM: Domain-Aware Symbolic Regression with LLMs for Kinetic Model Discovery
DASyR-LLM adds model critique and candidate generation to symbolic regression, cutting search iterations in simulated chemistry studies without improving final fit.
DASyR-LLM alternates symbolic regression with an LLM that critiques candidate equations and proposes replacements using chemical knowledge. Across **4 in silico case studies**, it cut iterations to the ground-truth model by **41.7–79.3%** and directly proposed the correct structure in over half of guided runs.
Builders of research agents should separate numerical search from domain critique, then feed the strongest candidates back into both. The ablations suggest a smaller LLM can retain much of the discovery efficiency, so model scale need not be the first lever.
DASyR-LLM alternates symbolic regression with an LLM that critiques candidate equations and proposes replacements using chemical knowledge. Across **4 in silico case studies**, it cut iterations to the ground-truth model by **41.7–79.3%** and directly proposed the correct structure in over half of guided runs. Builders of research agents should separate numerical search from domain critique, then feed the strongest candidates back into both. The ablations suggest a smaller LLM can retain much of the discovery efficiency, so model scale need not be the first lever. Both approaches reached **R² above 0.98** on independent validation, so the reported advantage is search efficiency rather than predictive quality. All studies were simulated; reduced wet-lab effort is an extrapolation, not a demonstrated outcome.
DASyR-LLM provides measured evidence that an LLM can accelerate a deterministic scientific search loop by critiquing and replacing candidate equations without improving final predictive fit. It reinforces architectures that reserve calculation and validation for specialized systems, while narrowing the claimed benefit to simulated search efficiency rather than real experimental savings.