SMEFT operator space without enumeration: a reinforcement-learning agent finds 26 low-overlap, high-quality solutions among 912 operators
Synopsis
With 912 baryon-, lepton- and lepton-flavour-conserving SMEFT operators, enumerating multi-operator combinations becomes a combinatorial explosion. ReinforceSMEFT casts operator-by-operator assembly as sequential decision-making: an actor–critic policy with about 40K trainable parameters learns which operator to add in an environment of 132 observables, 24 of which deviate from the Standard Model by more than 3σ, and is rewarded for solutions that overlap little with those already found. Against random-subset and random-sequential searches, the policy builds multi-operator theories more efficiently; combining ten independent training runs yields 26 high-quality solutions with Jaccard indices below 0.1.
Interpretation
Indirect searches for new physics no longer have to rely on physics intuition to narrow the operator list; a learned policy assembles multi-operator theories step by step in SMEFT operator space. Conventional phenomenological analyses use a specific ultraviolet theory, symmetry argument or simplified model to restrict candidate operators, covering only a small fraction of the full operator space, while exhaustive enumeration of all combinations is computationally prohibitive. On a benchmark environment of 132 observables, 24 of which deviate from the Standard Model by more than 3σ, the policy is compared with random-subset and random-sequential searches at comparable computational budgets and navigates this combinatorial space more efficiently.
The agent proposes one operator per step and keeps it only if the global fit improves, so it learns whether an operator is useful given the operators already present. A single operator can affect several observables at once, and operator mixing through renormalization-group evolution and matching means a multi-operator hypothesis cannot be judged from the performance of its members in isolation. The actor–critic network takes operator embeddings and observable pulls as input and relates them with multi-head cross-attention; all roughly 40K trainable parameters train on standard CPU resources, and the dependence of observables on Wilson coefficients is precomputed as semi-analytic expressions so each proposal does not rerun the renormalization-group evolution.
The search targets a set of structurally distinct high-quality hypotheses rather than a single best-fit theory. A policy optimized only for fit quality keeps returning to the neighbourhood of an already discovered high-quality operator set, producing near-duplicates that differ by one or a few operators. A diversity term based on Jaccard similarity is added to the reward, and only solutions passing the quality threshold enter the archive; combining ten independently trained policies identifies 26 solutions with Jaccard indices below 0.1 and Δχ² relative to the Standard Model of at least 100.
Perspective
The demonstrated setting is a hypothetical dataset of 132 observables spanning electroweak precision observables, K- and B-meson processes, τ decays, leptonic magnetic moments, Higgs observables and scattering measurements, with 24 measurements set to deviate from the Standard Model by more than 3σ; the operator space is restricted to 912 SMEFT operators conserving baryon number, lepton number and lepton flavour, and observables retain only linear dependence on the Wilson coefficients. The authors propose applying the framework to larger observable sets and updated experimental measurements and studying how policies transfer between different discrepancy patterns.
The benchmark dataset is called hypothetical because it mixes real and fictitious measurements, and the real measurements it includes are not necessarily up to date; family counts depend mildly on discovery order because membership is defined only relative to family representatives and is not transitive; the mean or maximum fit quality of an individual training batch does not rise monotonically, which the authors attribute to continued exploration rather than loss of learned solutions.
