OpenCompass's LLM adaptation of the SRBench symbolic-regression project: given numeric input-output samples, a model must output a closed-form formula, scored by fit quality and symbolic equivalence.
unassessed
| Category | math |
|---|---|
| Subcategory | symbolic regression from numeric data (LLM equation-discovery adaptation) |
| Page status | active |
| Metric | RMSE / NMSE / R² plus symbolic-equivalence match |
| Direction | lower_is_better |
| Dataset licence | GPL-3.0 |
| Publisher | University of Pennsylvania (Cava Lab) and collaborators; OpenCompass LLM adaptation by Shanghai AI Laboratory / open-compass |
SRBench (as run by OpenCompass) gives a model a table of numeric input-output samples drawn from a known physics equation and asks it to infer a closed-form symbolic formula using a restricted operator set (+, -, *, /, exp, sqrt, sin, arcsin and constants). This adapts the Feynman Symbolic Regression Database, the ground-truth physics-equation portion of the original SRBench project, into a prompted LLM task: the model sees example (x, y) pairs and must recover the underlying formula well enough to predict held-out points, rather than being told the equation's functional form.
Free-form text generation: the model reads two prompts describing the input variables and sample values, then must output only the inferred formula string. OpenCompass's config draws 300 random samples per problem and splits them roughly 97%/3% for the model to fit against and be evaluated on. No retrieval or tool use is part of the default config (ZeroRetriever, GenInferencer).
No model card in ModelSpec reports this benchmark yet.