Skip to main content
Back to timeline
arXivSource publication:

automodel turns model discovery into a skill: general coding agents find interpretable models across five sciences

Synopsis

Model discovery has relied either on manual design or on search pipelines built specifically for each problem, with structural choices and operator combinations fixed in advance. automodel packages the procedure of building an evaluator, fitting a baseline, iterating on structural hypotheses and running an external final check as a portable agent skill that general coding agents such as Claude Code and Codex execute. Across five case studies in stellar mass functions, cardiovascular modeling, precision agriculture, tire dynamics and traffic flow, the agents found new or improved models; on the stellar mass function, A5 lowered description length relative to the Bernardi function on both the Sérsic and cmodel datasets while using 19 operators against 23.

Source-provided article image: automodel: A Skill for Agentic Model Discovery
arXiv

Interpretation

Model discovery can now run as a portable agent skill executed by a general coding agent, rather than as a search pipeline built separately for each problem. Symbolic regression, neural architecture search and AutoML depend on dedicated search algorithms whose structural choices and allowed combinations the modeler must pre-specify and adapt to the data; systems such as AlphaEvolve need dedicated harness software to coordinate model calls, candidate generation and evaluation. Five case studies span astrophysics, cardiovascular dynamics, precision agriculture, traffic flow and tire dynamics, with model representations ranging from one-line analytical expressions to equation modifications inside scientific simulators, all carried out by commercial coding agents guided by automodel.

The skill splits discovery into setup, baseline, iterate and finalize phases, with the agent and researcher jointly fixing objectives, constraints and evaluation. The iterate phase is a nested loop: an orchestrating agent dispatches parallel sub-agents to non-overlapping search directions, and the sub-agents repeatedly change model structure, fit free parameters and record results before the orchestrator reviews them and issues updated directions. In the stellar mass function case, the agent ran Claude Code (Opus 4.8) for three meta-iterations of four parallel agents with four sequential structural edits each, and the evaluator reproduced the published description lengths, making candidates directly comparable with literature numbers.

On the stellar mass function, the search returned seven functions that improve on the best exhaustive symbolic regression result, and A5 beats the Bernardi function in description length on both datasets. Exhaustive symbolic regression is computationally limited to roughly complexity 10, while the Saunders and Bernardi functions contain 21 and 23 operators, so the exhaustive method cannot reach the complexity of those literature forms. A5 uses 19 operators against Bernardi's 23, lowers description length on both datasets and raises the likelihood; A5 also passes three physicality checks: positivity, decrease with mass, and an integral below the cosmological baryonic mass budget.

Perspective

The demonstrated scope is five concrete scientific modeling problems: fitting a galaxy stellar mass function, modifying CircAdapt cardiovascular equations, finding nitrogen-response skeletons for winter wheat, an analytical lateral tire-force model, and corrections to LWR traffic fundamental diagrams; readers should treat it as a researcher-led, interactive discovery workflow operating on existing code and data. The authors note that the comparison against the same coding agent without the skill was a preliminary test in the astrophysics case only, and that more systematic control tests and detailed comparisons with traditional search algorithms remain to be made.

The authors state that the discoveries are hypotheses whose mechanistic interpretation needs independent validation; in the cardiovascular case the held-out patient's improvement was not uniform and is read as a modest advantage rather than population-level generalization; the eight-parameter precision-agriculture curves have weak site-level parameter identifiability and should not be interpreted independently across sites; in the tire model some coefficients approached optimization bounds and should be read as components of the predictive law rather than independently identifiable physical quantities.