betaFeedback

WorksComputer Science

A Reinforcement Learning Method to Navigate SMEFT Operator Space

Abstract

In particle physics, discrepancies between experimental measurements and Standard Model (SM) predictions are commonly interpreted within new theoretical models. In this context, the Standard Model Effective Field Theory (SMEFT) provides a general framework encompassing a broad class of models. However, exploring the enormous space of possible SMEFT operators rapidly becomes a challenging combinatorial problem: many operators can contribute simultaneously to different observables, their effects are correlated, and distinct operator combinations can provide comparably good descriptions of the data. In this work, we formulate the exploration of the SMEFT operator space as a sequential decision-making problem using reinforcement learning (RL). Our RL agent employs a lightweight actor-critic architecture with approximately 40K trainable parameters in total. Rather than searching for a single best-fit theory, our method is designed to discover a diverse set of high-quality theory hypotheses. We test the method in a complex phenomenological environment comprising 132 observables across several sectors of particle physics, including 24 measurements that deviate from their SM predictions by more than 1.5σ1.5σ. We show that the learned RL policy efficiently constructs multi-operator theories that substantially improve the global fit. Comparisons with independent random-subset and random-sequential searches demonstrate the ability of the RL policy to navigate this large combinatorial space. Our results establish RL as a systematic framework for exploring diverse patterns of new physics (NP) interactions within SMEFT.

Introduction

The SM provides a successful description of particle physics over a wide range of energy scales. Nevertheless, it cannot account for several important observations such as neutrino masses, dark matter, and the baryon asymmetry of the universe. This motivates the presence of physics beyond the SM (BSM). At the same time, direct searches at the Large Hadron Collider (LHC) have so far revealed no conclusive evidence for such NP. If the scale of NP lies much above the energies directly accessible to current colliders, its effects may instead first appear indirectly through deviations of low-energy observables from their SM predictions, i.e., anomalies.

Effective field theories (EFTs) provide a model-independent framework for interpreting such discrepancies. Above the electroweak scale, the effects of heavy NP can be parameterized within the Standard Model Effective Field Theory (SMEFT) Buchmuller and Wyler [1986], Grzadkowski et al. [2010](for the reviews, see Brivio and Trott [2019], Isidori et al. [2024], Aebischer et al. [2026]), which contains only SM fields and respects the SU(3)C×SU(2)L×U(1)YSU(3)_{C} \times SU(2)_{L} \times U(1)_{Y} gauge symmetry. Below the electroweak scale, the appropriate description is the Weak Effective Theory (WET), obtained after integrating out the heavy SM degrees of freedom and respecting SU(3)C×U(1)emSU(3)_{C} \times U(1)_{\mathrm{em}}.

Statement of the problem: A central problem in indirect searches for NP is the following: given a collection of experimental measurements, which combinations of SMEFT operators can improve the agreement with the data while remaining compatible with the other experimental constraints? This bottom-up problem is an important intermediate step toward the eventual top-down question of identifying UV theories capable of generating the preferred SMEFT operators.

However, the difficulty arises from the enormous combinatorial space of the SMEFT operators. At dimension six, the Warsaw basis Grzadkowski et al. [ [8] contains 59 baryon and lepton number conserving operators, corresponding to 2499 independent operators once flavour indices are taken into account. Even when only a subset of these operators is relevant to a particular low-energy dataset, the number of possible multi-operator hypotheses grows rapidly. For NN candidate operators, the number of possible theories containing up to kk operators is

Ncomb=∑n=1k(Nn).(1)N_{\mathrm{comb}} = \sum_{n=1}^{k} \binom{N}{n}. \tag*{(1)}

Here NN is total number of SMEFT operators considered in a problem, and nn represents the number of operators in a given SMEFT hypothesis. In the present analysis, we work with 912 SMEFT operators that conserve the baryon number, lepton number, and lepton flavour. The challenge is therefore not simply to determine the continuous values of a set of WCs, but first to determine which operators should be present. This introduces a discrete combinatorial search problem on top of the usual continuous parameter estimation.

Importantly, SMEFT operators cannot generally be considered independent because a single operator may affect multiple observables. These correlations become even richer after renormalization-group (RG) evolution and matching between SMEFT and WET Aebischer et al. [ [2]. Operator mixing due to RG running implies that an operator introduced to improve one observable may simultaneously modify several others. In order to interpret experimental measurements within SMEFT, therefore it requires understanding of non-trivial correlations among different directions in the operator space.

These correlations are particularly important for multi-operator SMEFT hypotheses. An operator that improves the description of one discrepant observable may worsen agreement with another measurement. A second operator can then compensate for this deterioration while preserving, or further improving, the original effect. Subsequent operators can similarly repair constraints or address additional discrepancies. Consequently, the quality of a multi-operator hypothesis cannot in general be inferred from the performance of its constituent operators in isolation. Viable solutions can instead emerge through correlated combination of operators.

Conventional phenomenological analyses address this complexity using physics intuition to restrict the set of operators under consideration, often guided by a particular UV theory, symmetry argument, or simplified model. Such approaches are very useful for testing specific hypotheses, but necessarily explore only a small fraction of the complete SMEFT operator space. At the opposite extreme, exhaustive enumeration of all possible operator combinations quickly becomes computationally prohibitive. In this work, we formulate the search for viable SMEFT operators as a sequential decision making problem using an “actor-critic” RL algorithm Sutton et al. [ [20], Sutton and Barto [ [19] as introduced below.

Reinforcement learning method: We design an RL policy using a transformer Vaswani et al. [ [21] based neural-network (NN) which learns to propose viable SMEFT operators combinations through repeated interactions with phenomenological environment. So instead of evaluating every SMEFT operator combination independently, an RL policy can construct a multi-operator hypothesis sequentially, observe how the global fit to experimental data changes, and use this feedback to adjusts the NN weights to learn.

An important feature of this formulation is that the policy can learn conditional operator preferences. If the addition of one operator repeatedly produces a characteristic pattern of improvements and deterioration among observables, the policy can learn which subsequent operators are effective in that particular state. The RL agent2 therefore does not merely learn a global ranking of individual operators, it learns how useful an operator is conditional on the presence of other operators.

We stress that our objective is broader than identifying a single best-fit theory. Different combinations of SMEFT operators can provide comparably good descriptions of the same data, and identifying this degeneracy is itself important for model building. We therefore formulate the task as a quality–diversity search: the agent should discover high-quality solutions while continuing to explore structurally different regions of the SMEFT operator space. To achieve this, we supplement the fit-quality reward Kumar et al. [ [12] with a novelty-based reward function that discourages repeated sampling of closely related operator sets. The details of RL formulation are given in the next section.

Related efforts have recently explored the use of machine learning (ML) and artificial intelligence (AI) techniques for automated theory exploration in particle physics. For example, Ref. Baretz et al. [ [5] developed the Autonomous Model Builder (AMBer), which employs RL to construct viable neutrino-flavor models by selecting symmetry groups, particle content, and representation assignments. Ref. Alexander et al. [ [4] proposed the autonomous discovery of particle physics theories directly from experimental data using RL and symbolic-regression techniques. Ref. Hammad and Sanz [ [9] explored a complementary language-model-based approach in which a domain-adapted LLM generates hypotheses for sparse SMEFT analyses, while Ref. Saad [ [16] developed an LLM-assisted framework for BSM model building, in which an LLM provides a natural-language interface to a deterministic symbolic backend that automatically constructs and analyzes BSM models from user-specified field content.

These studies illustrate the growing interest in using ML to navigate large theory space and facilitate traditionally human-guided aspects of model building in particle physics.

Code availability: The method presented in this work builds upon my open-source GitHub implementation ReinforcePhysics. This RL framework includes implementations based on both the “actor–critic" algorithm and Group Relative Policy Optimization (GRPO)Shao et al. [ [17]; in the present work, we focus on the former. To facilitate reproducibility and broader use, we have packaged the complete methodology into the pip-installable Python package, which is publicly available and will be actively maintained in the ReinforceSMEFT GitHub repository.

Reinforcement Learning Formulation

In what follows, we provide details of our RL formulation of the SMEFT problem and its implementation. Starting from the SM, the RL agent proposes one SMEFT operator at each time step tt. The Wilson coefficients (WCs) of the currently proposed operators are optimized by minimizing the global χ2\chi^2, and the resulting improvement in the fit along with diversity criteria provides the primary feedback to the agent in terms of rewards. The correspondence between the RL and the SMEFT problem is, therefore:

  • Action ata_t: proposing an SMEFT operator;

  • Reward rtr_t: improvement in fit quality and solution diversity;3

  • State sts_t contains the features – current observable pulls, χ2\chi^2, selected operators, previously evaluated operators, and trajectory step tt;

  • Policy πθ(at∣st)\pi_\theta(a_t|s_t): is the rule of decision to be learned that assigns probabilities to candidate SMEFT operators or actions ata_t, θ\theta are learnable parameters;

  • Environment: observable predictions within the SMEFT, input experimental measurements, global χ2\chi^2 minimizer.

The pull of the iith observable is defined as pi=(Oiexp−OiSM)/(σiexp)2+(σiSM)2p_i = (O_i^{\mathrm{exp}} - O_i^{\mathrm{SM}})/\sqrt{(\sigma_i^{\mathrm{exp}})^2 + (\sigma_i^{\mathrm{SM}})^2}, where OiexpO_i^{\mathrm{exp}} and OiSMO_i^{\mathrm{SM}} denote the experimental measurement and the SM prediction, respectively, while σiexp\sigma_i^{\mathrm{exp}} and σiSM\sigma_i^{\mathrm{SM}} are their corresponding uncertainties. A positive (negative) pull indicates that the experimental value lies above (below) the SM prediction.

The policy sequentially selects SMEFT operators, thereby generating a trajectory of states

s0→O1s1→O2s2→O3⋯→OTsT(2)s_0 \xrightarrow{O_1} s_1 \xrightarrow{O_2} s_2 \xrightarrow{O_3} \cdots\xrightarrow{O_T} s_T \tag*{(2)}

where OtO_t denotes the operator selected at step tt, and sts_t represents the state of the environment after the first tt operators have been selected. Initially, a proposed operator is temporarily added to the currently accepted operator set and the corresponding χ2\chi^2 is minimized. The operator is retained if it improves the fit by more than Δχmin⁡2=2\Delta\chi^2_{\min}=2 and is rejected otherwise. In either case, it is marked as evaluated so that the same operator is not proposed again during the trajectory. The process is repeated for at most T=10T=10 steps (i.e., the trajectory length).

At each step, the state encodes information about the current quality of the global fit, together with the operators selected and evaluated up to that point. Based on the current state, the policy assigns a probability to each available operator

Ot+1∼πθ(O∣st),(3)O_{t+1} \sim\pi_\theta(O \mid s_t), \tag*{(3)}

and the selected operator determines the transition from sts_t to st+1s_{t+1}. During training, the policy learns to assign larger probabilities to operator selections that lead to larger cumulative future rewards. Consequently, the agent is not restricted to selecting operators that provide the largest immediate improvement in the fit; it can learn sequences in which an operator produces only a modest initial improvement but enables subsequent operators to achieve a substantially better global fit.

During the RL training itself we save all possible operator hypotheses (here a single hypothesis is defined by accepted operators in a trajectory in TT time steps) sampled from the policy and post-training we recover all viable hypotheses based on some criteria; such as minimum Δχ2\Delta\chi^2 or number of operators. This approach is similar to Deep Symbolic Regression Petersen et al. [ [15] which discovers mathematical expressions describing a dataset during the RL training.

Actor–critic learning

The “actor–critic" algorithm is used to train the policy. The basic idea of the actor–critic algorithm is straightforward: actor is simply the policy πθ(at∣st)\pi_\theta(a_t|s_t), which we parameterize by a small transformer-based NN. It determines the probability of selecting action ata_t (or an SMEFT operator OtO_t) in a given state sts_t. The critic estimates the expected future return from the current state and thereby provides a baseline for the policy-gradient update.

Here, the return is defined by the cumulative discounted reward obtained from the current state onward when following the current policy πθ\pi_\theta. For a sampled trajectory, the return at step tt is

Gt=∑k=0T−tγkrt+k,G_t = \sum_{k=0}^{T-t} \gamma^k r_{t+k},

where rt+kr_{t+k} is the reward received at a future step and γ\gamma is the discount factor.

The actor, with learnable parameters θ\theta, is optimized using the loss

Lactor(θ)=−1N∑tAtlog⁡πθ(at∣st).(4)\mathcal{L}_{\mathrm{actor}}(\theta) = -\frac{1}{N}\sum_t A_t \log\pi_\theta(a_t|s_t). \tag*{(4)}

Here, AtA_t is the advantage, which quantifies whether the sampled action performs better or worse than the expected behavior of the policy from that state. In terms of the so-called “true value function" Vπ(st)V^\pi(s_t), the advantage is defined as

At=Gt−Vπ(st).(5)A_t = G_t - V^\pi(s_t). \tag*{(5)}

However, since the true value function Vπ(st)V^\pi(s_t) is not known, we train a separate NN, the critic, to approximate it by Vϕ(st)V_\phi(s_t), where ϕ\phi denotes the learnable critic parameters.

For each batch of sampled trajectories, the critic is trained using the objective

Lcritic(ϕ)=1N∑t[Vϕ(st)−Gt]2.(6)\mathcal{L}_{\mathrm{critic}}(\phi) = \frac{1}{N}\sum_t [V_\phi(s_t)-G_t]^2. \tag*{(6)}

Here, NN denotes the total number of sampled state–action steps entering the corresponding training update. The estimated value function Vϕ(st)V_\phi(s_t) is then used to construct the advantage: At≈Gt−Vϕ(st)A_t \approx G_t - V_\phi(s_t), which enters the actor loss in Eq. (4).

The total training objective combines the actor and critic losses

Ltotal=Lactor+λVLcritic(7)\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{actor}} + \lambda_V \mathcal{L}_{\mathrm{critic}} \tag*{(7)}

where λV=0.5\lambda_V = 0.5 controls the relative contribution of the critic loss.

Training proceeds iteratively by sampling batches of trajectories from the current policy πθ\pi_\theta and using them to update the network parameters. For the actor, the gradient ∇θLactor\nabla_\theta\mathcal{L}_{\mathrm{actor}} increases the probability of actions with positive advantage and suppresses those with negative advantage. Importantly, AtA_t is treated as a fixed target when computing the actor gradient. Consequently, for separate actor and critic networks, the critic loss does not contribute directly to the gradient with respect to θ\theta; instead, the critic influences the actor update through the state-dependent baseline Vϕ(st)V_\phi(s_t) used to construct the advantage.

In our implementation, Lactor\mathcal{L}_{\mathrm{actor}} and Lcritic\mathcal{L}_{\mathrm{critic}} are combined into the single objective in Eq. (7) and optimized in one backward pass. An RL algorithm that does not employ a critic and instead uses Monte Carlo returns directly in the policy-gradient estimator is commonly known as REINFORCE Williams [ [23]. The principal role of the critic is therefore to reduce the variance of the policy-gradient estimate by providing a state-dependent baseline, while the actor learns which operator selections lead to larger-than-expected returns.

State representation and NN architecture

At each step tt, the RL agent observes the Markov state

st=(pt,χt2,mtactive,mteval,t)(8)s_t = \left(\mathbf{p}_t, \chi_t^2, \mathbf{m}_t^{\mathrm{active}}, \mathbf{m}_t^{\mathrm{eval}}, t\right) \tag*{(8)}

where pt=(p1,t,…,pNobs,t)\mathbf{p}_t = (p_{1,t}, \ldots, p_{N_{\mathrm{obs}},t}) contains the pulls which quantify the deviations of the observables w.r.t. the SM, χt2\chi_t^2 quantifies the overall agreement with the data, mtactive\mathbf{m}_t^{\mathrm{active}} specifies the operators currently included in the SMEFT scenario, and mteval\mathbf{m}_t^{\mathrm{eval}} records the operators that have already been evaluated during the trajectory. The latter prevents the duplicate operator from being proposed repeatedly. The step number tt provides information about the position within the trajectory. Both the actor and critic receive this same state representation.

As we have a fixed number of SMEFT operators in the problem, each candidate operator OiO_i can be represented by a trainable embedding

Oi⟶eiop∈Rd.(9)O_i \longrightarrow\mathbf{e}_i^{\mathrm{op}} \in\mathbb{R}^{d}. \tag*{(9)}

In our implementation, the latent dimension (dd) is 32. Each observable pull is projected independently into the same latent space

pj,t⟶ej,tp=Wppj,t+bp∈Rd.(10)\mathbf{p}_{j,t} \longrightarrow\mathbf{e}_{j,t}^{p} = W_p\mathbf{p}_{j,t} + \mathbf{b}_p \in\mathbb{R}^{d}. \tag*{(10)}

Here, WW and b\mathbf{b} denote a subset of the trainable parameters of the NN. To relate the candidate operators to the current pattern of deviations in the observable space, we employ multi-head cross-attention, taking the operator embedding as queries and the pull embedding as keys and values, i.e. Q=EopQ = E^{\mathrm{op}}, K=Ep(st)K = E^p(s_t) and V=Ep(st)V = E^p(s_t).

For a single attention head, this operation takes the standard form Vaswani et al. [ [21]

Attn⁡(Q,K,V)=softmax⁡(QKTdh)V(11)\operatorname{Attn}(Q,K,V) = \operatorname{softmax}\left(\frac{QK^{T}}{\sqrt{d_h}}\right)V \tag*{(11)}

here, we use four attention heads, i.e., h=4h = 4.

The resulting vector ci,t∈Rd\mathbf{c}_{i,t} \in\mathbb{R}^{d} – the ii-th row of Attn – provides a state-dependent representation of the observable information relevant to candidate operator OiO_i. In particular, the attention mechanism allows different operators to focus on different patterns of observable pulls.

The remaining components of the state are encoded separately. The binary active and evaluated operators are mapped to trainable embedding

mi,tactive⟶ei,tactive,mi,teval⟶ei,teval(12)m_{i,t}^{\mathrm{active}} \longrightarrow\mathbf{e}_{i,t}^{\mathrm{active}}, \qquad m_{i,t}^{\mathrm{eval}} \longrightarrow\mathbf{e}_{i,t}^{\mathrm{eval}} \tag*{(12)}

while the normalized global quantities (χt2/χref2,t/Tmax⁡)(\chi_t^2/\chi_{\mathrm{ref}}^2, t/T_{\max}) are passed through a small multi-layer perceptron to obtain a global representation gt∈Rd\mathbf{g}_t \in\mathbb{R}^{d}.

The complete state-dependent representation of operator OiO_i is then constructed by concatenating these components

hi,t=[eiop; ci,t; ei,tactive; ei,teval; gt]∈R5d.(13)\mathbf{h}_{i,t} = \left[\mathbf{e}_i^{\mathrm{op}};\,\mathbf{c}_{i,t};\,\mathbf{e}_{i,t}^{\mathrm{active}};\,\mathbf{e}_{i,t}^{\mathrm{eval}};\,\mathbf{g}_t\right] \in\mathbb{R}^{5d}. \tag*{(13)}

For d=32d = 32, each candidate operator is represented by a 160-dim vector. This representation encodes the operator identity together with its attention-weighted relation to the current observable pulls, whether it has already been selected or evaluated, and information about the current global state of the trajectory. From a physics perspective, this representation allows the relevance of each operator to depend dynamically on the pattern of deviations of observables from the SM and on the operators already selected, thereby enabling the policy to learn correlations and complementary effects among different operators.

The actor maps each representation hi,th_{i,t} to a scalar logit zi,tz_{i,t}

zi,t=fθ(hi,t),(14)z_{i,t} = f_{\theta}(h_{i,t}), \tag*{(14)}

from which the policy is obtained as

πθ(Oi∣st)=exp⁡(zi,t)∑j∈Atexp⁡(zj,t),(15)\pi_{\theta}(O_i \mid s_t) = \frac{\exp(z_{i,t})}{\sum_{j \in\mathcal{A}_t} \exp(z_{j,t})}, \tag*{(15)}

where At\mathcal{A}_t denotes the set of operators that have not yet been evaluated. Previously evaluated operators are explicitly masked and hence have zero probability of being selected again.

As the role of critic is to determine how promising the current state is, irrespective of which operator we choose, it only needs one global representation of the current state, which is obtained using

hˉt=1Nop∑i=1Nophi,t.(16)\bar{h}_t = \frac{1}{N_{\mathrm{op}}} \sum_{i=1}^{N_{\mathrm{op}}} h_{i,t}. \tag*{(16)}

Here NopN_{\mathrm{op}} is number of sampled operators at time step tt. After concatenating with the global quantities χt2\chi_t^2, tt, i.e., gt\mathbf{g}_t the critic then predicts the scalar state value

Vϕ(st)=fϕ([hˉt;gt]).(17)V_{\phi}(s_t) = f_{\phi}\left([\bar{h}_t; \mathbf{g}_t]\right). \tag*{(17)}

Thus, the actor learns which operator should be selected next, whereas the critic estimates the expected quality of the current state.

Crucially, the operator embedding, pull projection, attention mechanism, state embedding, actor, and critic are trained jointly through the RL objective. Consequently, the network learns a state-dependent representation of the SMEFT operator space: the preference for a given operator can change as the pattern of pulls and the set of previously selected operators evolve along a trajectory.

The NN architecture and RL training framework are implemented using PyTorch [2019]. For the full action space of 912 SMEFT operators and an embedding dimension d=32d = 32, the actor-critic NNs contain approximately ∼\sim40K trainable parameters and is sufficiently lightweight to be trained efficiently using standard CPU resources, without requiring GPU acceleration. In practice, the computational cost is dominated by repeated evaluation and optimization of the physics environment (which is mainly the global χ2\chi^2-minimization) rather than by the NN operations.

Reward function

In our setup, at each interaction, the environment provides feedback to the policy through a scalar reward signal. In addition to improving the quality of the fit, our goal is to encourage the policy to explore distinct regions of the SMEFT operator space. Accordingly, the reward function consists of two components: a physics reward associated with the improvement in fit quality and a diversity reward that promotes the discovery of distinct operator combinations. The diversity reward is evaluated and added at the end of each trajectory. The total reward is therefore defined as

rt=rtfit+λdiv(1−JmaxA)I(t=T).(18)r_t = r_t^{\mathrm{fit}} + \lambda_{\mathrm{div}}(1 - J_{\mathrm{max}}^{A})\mathbb{I}(t = T). \tag*{(18)}

Here, parameter λdiv\lambda_{\mathrm{div}} controls the relative importance of diversity compared to improvement of the physics fit and I(t=T)\mathbb{I}(t = T) is an indicator function that equals 1 only at the final step t=Tt = T, and 0 at all earlier steps, ensuring that the novelty reward is added only at the end of the trajectory.

The fit reward, first term of (18), is proportional to the improvements in the fit, given by

rtfit=Δχt2−αstep−λdeadI(Δχt2<Δχmin2)+λsolI(χt2<χsol2).(19)r_t^{\mathrm{fit}} = \Delta\chi_t^2 - \alpha_{\mathrm{step}} - \lambda_{\mathrm{dead}}\mathbb{I}(\Delta\chi_t^2 < \Delta\chi_{\mathrm{min}}^2) + \lambda_{\mathrm{sol}}\mathbb{I}(\chi_t^2 < \chi_{\mathrm{sol}}^2). \tag*{(19)}

Here, Δχt2=χt−12−χt2\Delta\chi_t^2=\chi_{t-1}^2-\chi_t^2 denotes the reduction in the global χ2\chi^2 obtained after the action at step tt, such that a positive value corresponds to an improvement of the fit. The parameter αstep\alpha_{\mathrm{step}} provides a small cost for each operator-selection step, favoring solutions with a small number of operators. The third term penalizes unproductive actions, with λdead\lambda_{\mathrm{dead}} denoting the dead-step penalty and Δχmin⁡2\Delta\chi_{\min}^2 the minimum improvement required for an action to be considered useful. The final term in (19) provides an additional reward λsol\lambda_{\mathrm{sol}} when the resulting theory reaches the target fit quality χsol2\chi_{\mathrm{sol}}^2.

For the diversity reward given by the second term of (18), the similarity between two operator sets is quantified by the Jaccard index: J(ST,S′)=∣ST∩S′∣/∣ST∪S′∣J(S_T,S')=\lvert S_T\cap S'\rvert/\lvert S_T\cup S'\rvert. The largest similarity of current trajectory STS_T to any other set already contained in the archive A\mathcal{A} is given by Jmax⁡A=max⁡S′∈AJ(ST,S′)J_{\max}^{\mathcal{A}}=\max_{S'\in\mathcal{A}}J(S_T,S'), here, A\mathcal{A} is an archive of previously discovered high-quality solutions. Only solutions satisfying the chosen quality requirement, Δχ2>Δχarchive2\Delta\chi^2>\Delta\chi_{\mathrm{archive}}^2, are added to this diversity archive. Consequently, a solution identical or very similar to an archived solution receives little diversity reward, whereas a solution occupying a previously unexplored region of operator space receives a larger reward.

RL Environment

The SMEFT operators are sampled sequentially using the current policy πθ\pi_\theta. After each operator is sampled in a trajectory, our framework invokes an internal χ2\chi^2-minimization procedure to evaluate the resulting fit and determine the corresponding reward. Since training the RL policy requires a large number of such interactions with the environment, this procedure must be repeated many times. A direct evaluation of the χ2\chi^2 function would require performing the full RG evolution within SMEFT and WET, followed by the computation of the relevant observables, for every proposed set of operators. Combined with the repeated χ2\chi^2 minimization, this would render the RL training prohibitively expensive.

To overcome this computational bottleneck, we precompute semi-analytic expressions for the observables, retaining their linear dependence on the SMEFT WCs defined at the NP scale. The computationally expensive RG evolution and observable calculations are therefore performed only once during this preprocessing stage. The resulting expressions can subsequently be evaluated very efficiently within the RL environment, enabling rapid computation of the χ2\chi^2 function and the corresponding rewards throughout training.

Semianalytic χ2\chi^2 function

At dimension-six level in SMEFT, the theoretical prediction for an observable aa can be written as

Oth,a({Ci})=OSM,a+∑iαaiCi(Λ)+∑j,kj≤kβa,jkCj(Λ)Ck(Λ)(20)\mathcal{O}_{\mathrm{th},a}(\{C_i\})=\mathcal{O}_{\mathrm{SM},a}+\sum_i\alpha_{ai}C_i(\Lambda)+\sum_{\substack{j,k\\j\leq k}}\beta_{a,jk}C_j(\Lambda)C_k(\Lambda) \tag*{(20)}

where Ci(Λ)C_i(\Lambda) are the WCs defined at the NP scale Λ\Lambda. The coefficients αai\alpha_{ai} and βa,jk\beta_{a,jk} encode, respectively, the linear and quadratic dependence of observable aa on the SMEFT WCs, including the effects of RG evolution and matching at the electroweak scale. We assume a CP-conserving SMEFT scenario, such that the WCs are real. In the present analysis, we retain only the linear contribution proportional to αai\alpha_{ai} and construct the corresponding semi-analytic expression for each observable included in the global fit. We emphasize, however, that the RL method itself is not restricted to the linear approximation and can also be applied when quadratic terms are retained.

The χ2\chi^2 function is defined as

χ2({Ci})=∑a[Oth,a({Ci})−Oexp,a]2σth,a2+σexp,a2(21)\chi^2(\{C_i\})=\sum_a\frac{\left[\mathcal{O}_{\mathrm{th},a}(\{C_i\})-\mathcal{O}_{\mathrm{exp},a}\right]^2}{\sigma_{\mathrm{th},a}^2+\sigma_{\mathrm{exp},a}^2} \tag*{(21)}

where Oexp,a\mathcal{O}_{\mathrm{exp},a} denotes the experimental measurement, while σth,a\sigma_{\mathrm{th},a} and σexp,a\sigma_{\mathrm{exp},a} denote the corresponding theoretical and experimental uncertainties, respectively. To determine the semi-analytic coefficients, we use flavio Straub [2018] for the computation of observable predictions and wilson Aebischer et al. [2018a] for RG evolution.

The computationally expensive SMEFT calculations are therefore performed only once when constructing the semi-analytic expressions, while the subsequent χ2\chi^2 evaluations required repeatedly during RL training can be carried out efficiently using these precomputed expressions. We find that the RL training would be extremely slow without following this procedure.

Benchmark dataset

To test the RL method introduced in this work, we construct a sufficiently complex benchmark environment. For this purpose, we consider a hypothetical dataset4 consisting of 132 observables drawn from several sectors of particle physics, including electroweak precision observables, BB- and KK-meson processes, τ\tau decays, leptonic magnetic moments, Higgs observables, and scattering measurements. Within SMEFT, these sectors are not independent: correlations arise through gauge symmetry, RG evolution, matching, and dimension-six modifications of SM input parameters Aebischer et al. [ [2].

Among the 132 observables in our dataset, 24 differ from their chosen SM predictions by more than 1.5σ1.5\sigma, resulting in χSM2=273\chi_{\mathrm{SM}}^2 = 273. These observables are listed in Table 1. The remaining measurements act as constraints on the candidate operator sets.

ObservablePullHypothetical measurementSM prediction
ΔMs\Delta M_s−5.06-5.06(8.10±0.49)⋅10−12(8.10 \pm0.49)\cdot10^{-12}(1.14±0.046)⋅10−11(1.14 \pm0.046)\cdot10^{-11}
ϵ′/ϵ\epsilon'/\epsilon−4.53-4.53(−3.65±0.57)⋅10−3(-3.65 \pm0.57)\cdot10^{-3}(−2.76±55.57)⋅10−5(-2.76 \pm55.57)\cdot10^{-5}
BR(K+→πννˉ)\mathrm{BR}(K^+ \to\pi\nu\bar{\nu})+4.32+4.32(1.12±0.051)⋅10−10(1.12 \pm0.051)\cdot10^{-10}(8.14±0.50)⋅10−11(8.14 \pm0.50)\cdot10^{-11}
⟨P5′⟩(B0→K∗μμ)[4,6]\langle P'_5\rangle(B^0 \to K^*\mu\mu)^{[4,6]}−3.94-3.94−1.1607±0.0730-1.1607 \pm0.0730−0.7583±0.0721-0.7583 \pm0.0721
mWm_W+3.83+3.8380.3992±0.008480.3992 \pm0.008480.3551±0.007880.3551 \pm0.0078
Rτℓ(B→Dℓν)R_{\tau\ell}(B \to D\ell\nu)+3.79+3.790.3275±0.00660.3275 \pm0.00660.2950±0.00560.2950 \pm0.0056
AFB(Z→bbˉ)A_{\mathrm{FB}}(Z \to b\bar{b})−3.38-3.380.10239±0.000250.10239 \pm0.000250.10352±0.000220.10352 \pm0.00022
⟨dBR/dq2⟩(Bs→ϕμμ)[1,6]\langle d\mathrm{BR}/dq^2\rangle(B_s \to\phi\mu\mu)^{[1,6]}−3.36-3.36(2.4556±0.6031)⋅10−8(2.4556 \pm0.6031)\cdot10^{-8}(5.3556±0.6175)⋅10−8(5.3556 \pm0.6175)\cdot10^{-8}
⟨dR/dθ⟩(198.38,0.8,1.0)\langle dR/d\theta\rangle(198.38,0.8,1.0)−2.97-2.976.535±0.2366.535 \pm0.2367.2367.236
BR(W→τν)\mathrm{BR}(W \to\tau\nu)+2.56+2.560.11380±0.002100.11380 \pm0.002100.108417±0.0000490.108417 \pm0.000049
μtth(h→WW)\mu_{tth}(h \to WW)+2.29+2.295.00±1.755.00 \pm1.751.001.00
⟨dR/dθ⟩(189.09,0.8,1.0)\langle dR/d\theta\rangle(189.09,0.8,1.0)−2.20-2.205.762±0.2235.762 \pm0.2236.2536.253
BR(τ→μνν)\mathrm{BR}(\tau\to\mu\nu\nu)+2.06+2.060.17390±0.000400.17390 \pm0.000400.172894±0.0002820.172894 \pm0.000282
μZh(h→WW)\mu_{Zh}(h \to WW)+2.05+2.055.90±2.395.90 \pm2.391.001.00
⟨dR/dθ⟩(198.38,−0.6,−0.4)\langle dR/d\theta\rangle(198.38,-0.6,-0.4)+1.86+1.860.991±0.0840.991 \pm0.0840.8350.835
⟨dR/dθ⟩(198.38,0.6,0.8)\langle dR/d\theta\rangle(198.38,0.6,0.8)−1.77-1.774.100±0.1854.100 \pm0.1854.4284.428
⟨dR/dθ⟩(182.66,−1.0,−0.8)\langle dR/d\theta\rangle(182.66,-1.0,-0.8)−1.75-1.750.502±0.1140.502 \pm0.1140.7020.702
ϵK\epsilon_K+1.73+1.73(2.2280±0.0110)⋅10−3(2.2280 \pm0.0110)\cdot10^{-3}(1.9529±0.1583)⋅10−3(1.9529 \pm0.1583)\cdot10^{-3}
⟨dR/dθ⟩(198.38,−1.0,−0.8)\langle dR/d\theta\rangle(198.38,-1.0,-0.8)+1.73+1.730.679±0.0790.679 \pm0.0790.5420.542
⟨dR/dθ⟩(182.66,0.0,0.2)\langle dR/d\theta\rangle(182.66,0.0,0.2)+1.72+1.722.150±0.2442.150 \pm0.2441.7311.731
μWh(h→ττ)\mu_{Wh}(h \to\tau\tau)−1.71-1.71−1.40±1.40-1.40 \pm1.401.001.00
⟨dR/dθ⟩(205.92,−0.6,−0.4)\langle dR/d\theta\rangle(205.92,-0.6,-0.4)−1.70-1.700.653±0.0690.653 \pm0.0690.7700.770
R(e+e−→WW)(182.7)R(e^+e^- \to WW)(182.7)+1.64+1.641.036±0.0221.036 \pm0.0221.0001.000
μZh(h→bbˉ)\mu_{Zh}(h \to b\bar{b})−1.50-1.500.40±0.400.40 \pm0.401.001.00

Table 1. Anomalies (24 in total) included in the RL analysis, ordered by decreasing absolute SM pull, ∣pi∣|p_i|. The pull is defined as pi=(Oiexp−OiSM)/(σiexp)2+(σiSM)2p_i = (O_i^{\mathrm{exp}}-O_i^{\mathrm{SM}})/\sqrt{(\sigma_i^{\mathrm{exp}})^2+(\sigma_i^{\mathrm{SM}})^2}. The hypothetical experimental measurements and SM predictions are quoted with their corresponding experimental and theoretical uncertainties, respectively. The full dataset of 132 observables is available in the GitHub repository ReinforceSMEFT.

The sole purpose of this dataset is to provide a challenging benchmark containing a complex pattern of discrepancies with respect to the SM for the RL search algorithm. Our aim is to test whether the RL policy can efficiently navigate a large operator space in the presence of simultaneous discrepancies across different physical processes, while accounting for the fact that operators that improve the description of one subset of observables may also affect many others.

Training of the RL Policy

The objective of RL training is to learn a policy (or policies) through repeated interaction with the environment. The SMEFT operators that produce large improvements in the fit and satisfy the diversity criteria become more likely, whereas choices associated with poor trajectories are suppressed. As an example, for training, we use 300 batches, with 25 stochastic trajectories sampled per batch and at most T=10T = 10 proposals per trajectory. Since the RL search is sequential, in total, it involves 300×25×10=75K300 \times25 \times10 = 75\mathrm{K} calls to the internal χ2\chi^2-minimization loop. Without using the pre-computed semi-analytical expressions, it would be computationally very expensive. We use MINUIT James and Roos [1975] for the χ2\chi^2 minimization.

The actor and critic are optimized using AdamW with a learning rate of 10−310^{-3}. We use a discount factor γ=0.99\gamma= 0.99, critic-loss coefficient λV=0.5\lambda_V = 0.5, and gradient clipping at 0.5. An entropy term is included during training to prevent premature collapse of the policy and maintain exploration; its coefficient is annealed from 10−210^{-2} to 10−410^{-4}. The complete set of training hyperparameters and other details can be found on GitHub ReinforceSMEFT.

Figure 1 shows the evolution of fit quality during training. For each batch, we display the mean and maximum Δχ2\Delta\chi^2 obtained by the sampled trajectories, together with the best value discovered up to that point. The raw batch values fluctuate substantially because the policy remains stochastic and is explicitly encouraged to explore through both entropy regularization and diversity reward. Consequently, the mean or maximum Δχ2\Delta\chi^2 of an individual batch need not increase monotonically even after the policy has discovered high-quality regions of the operator space.

Figure 1. Learning curve of the RL policy for seed=42. For each training batch, the mean Δχ2\Delta\chi^2 relative to the SM is shown, together with the maximum Δχ2\Delta\chi^2 discovered up to that batch. The mean and maximum curves are smoothed using a 10-batch rolling average; the faint light green curve shows the corresponding unsmoothed values.

The cumulative best curve provides a complementary view of learning: once a high-quality solution is discovered, it remains part of a solution archive even while subsequent trajectories explore alternative regions. The late-time fluctuations should therefore not be interpreted as a loss of previously learned solutions, but rather as a consequence of continued exploration in a quality–diversity search.

Exploration of the SMEFT Operator Space

In this section, we analyze the viable SMEFT solutions5 discovered by the RL agent. To identify these solutions, post-training, we collect all candidate solutions sampled by the policy during training and rank them according to their Δχ2\Delta\chi^2 values w.r.t. the SM.

We stress that, a policy optimized solely for fit quality can repeatedly return to the neighborhood of a previously discovered high-quality operator set, generating many solutions that differ by only one or a few operators. Such solutions can be distinct as exact operator sets while representing essentially the same underlying neighborhood in operator space. Our objective is instead to discover both high-quality solutions and a broad range of structurally distinct operator combinations. The diversity term in Eq. (18) addresses this issue by rewarding solutions that have small Jaccard overlap with the existing archive of high-quality theories.

To quantify the resulting coverage of SMEFT hypotheses, we distinguish between exact solutions – corresponding to unique operator sets, and solution families – which group sufficiently similar operator sets according to their Jaccard similarity.

Exact solutions: Each candidate theory is specified by its set of active operators SS. For illustration, we consider a theory a solution if it satisfies Δχ2>100\Delta\chi^2 > 100 relative to the SM, and we count two solutions as the same exact solution if they contain identical operator sets, irrespective of operator order.

Solution families: To quantify how many structurally different solutions are found, we additionally group exact solutions into families using the Jaccard similarity. Solutions are processed in the order in which they are discovered, batch by batch. A new solution SS is compared with the stored family representatives {Rk}\{R_k\}. If J(S,Rk)>0.5J(S, R_k) > 0.5 for any kk, SS is assigned to an existing family and the number of families is unchanged. Otherwise SS becomes the founder of a new family and is stored as its representative. Thus the founder of a family is the first solution encountered that is not similar to any earlier representative, and is not necessarily the solution with the largest Δχ2\Delta\chi^2.

Because membership is defined only relative to the representative, it is not transitive, and the precise family count depends mildly on the discovery order. The same procedure and threshold are applied to the RL archive and to both random baselines as discussed below.

Comparison with random-search baselines

To benchmark the RL search, we consider two random-search baselines with comparable computational budgets. In the random-subset baseline, the number of operators in each subset is chosen randomly, and the corresponding operators are then sampled independently from the full operator space and evaluated directly. This provides a simple measure of how frequently high-quality hypotheses can be found through unguided sampling of the combinatorial operator space.

The second benchmark, the random-sequential baseline, follows the same sequential construction and acceptance rule as the RL agent, but selects each proposed operator uniformly at random. This provides a more direct algorithmic comparison because both methods construct theories sequentially and receive the same physics evaluation, but only the RL agent uses a learned state-dependent policy to select the next operator.

Figure 2 compares the cumulative discovery of solutions satisfying Δχ2>100\Delta\chi^{2} > 100 during the RL training. This threshold deliberately focuses the comparison on high-quality theories rather than on the total number of sampled configurations. The exact-solution count measures the number of unique operator sets discovered, whereas the family count provides a more stringent measure of operator diversity defined by J≤0.5J \le0.5. The comparison therefore probes two complementary aspects of the search: (i) whether the learned policy can concentrate on SMEFT operator space containing good fits and (ii) whether it can continue to discover distinct regions rather than repeatedly sampling small variations of the same operator set, e.g. by changing one SMEFT operator.

Figure 2. Cumulative number of high-quality SMEFT solutions discovered during the training as a function of batch. Only theories with Δχ2>100\Delta\chi^2 > 100 relative to the SM are retained. Exact operator sets and Jaccard families are shown for the RL agent and the random-search baselines. The random-sequential baseline follows the same step-wise construction and acceptance procedure as the RL agent but proposes operators uniformly at random.

These results show that the RL search fulfills our objective of not only maximizing Δχ2\Delta\chi^{2} but also efficiently exploring the high-quality regions of the SMEFT operator space.

Coverage across independent training seeds

Because RL training is stochastic, independent runs can converge toward different regions of the SMEFT operator space. However, this feature can be exploited by combining the solution archives obtained from multiple training seeds. Figure 3 shows the cumulative coverage obtained by combining ten independently trained RL policies. Exact operator sets typically grow more rapidly than the number of families because different runs can discover variations of the same underlying structures. The family curve therefore provides the more informative measure of whether additional independent training runs continue to probe genuinely different regions of operator space.

Figure 3. Cumulative coverage of the SMEFT solution space obtained by combining independent RL training runs. Only theories satisfying Δχ2>100\Delta\chi^{2} > 100 are included. The curves show the cumulative numbers of distinct exact operator sets and Jaccard families as additional training seeds are included.

In this figure, the continued growth of the family count indicates that additional runs still uncover new structures, whereas saturation would suggest that the dominant families accessible to the search have largely been identified. In total, we identified 26 solutions with Δχ2>100\Delta\chi^{2} > 100 and with J≤0.1J \le0.1, which are listed in Table 2 along with corresponding Δχ2\Delta\chi^2 values w.r.t. the SM hypothesis. The corresponding definitions of the SMEFT operators are provided in Table 3.

#WCsΔχ2\Delta\chi^2
1Clequ,3332(1)C_{lequ,3332}^{(1)}, Clequ,2232(3)C_{lequ,2232}^{(3)}, CHd,33C_{Hd,33}, CHq,33(3)C_{Hq,33}^{(3)}, CHu,33C_{Hu,33}, Cquqd,2311(1)C_{quqd,2311}^{(1)}, Cquqd,2323(1)C_{quqd,2323}^{(1)}142.9
2CdB,33C_{dB,33}, CdG,22C_{dG,22}, Cld,1112(1)C_{ld,1112}^{(1)}, Cledq,3333C_{ledq,3333}, Clequ,2213(3)C_{lequ,2213}^{(3)}, CHd,22C_{Hd,22}, CHq,2223(1)C_{Hq,2223}^{(1)}, Cqd,1321(8)C_{qd,1321}^{(8)}, Cqu,2323(1)C_{qu,2323}^{(1)}, Cquqd,1321(1)C_{quqd,1321}^{(1)}137.6
3CdW,22C_{dW,22}, CeW,11C_{eW,11}, Cld,1123(1)C_{ld,1123}^{(1)}, Cledq,3332C_{ledq,3332}, Clq,1123(1)C_{lq,1123}^{(1)}, CHd,2223C_{Hd,2223}, CHq,2223(1)C_{Hq,2223}^{(1)}, Cqd,2323(8)C_{qd,2323}^{(8)}, Cud,2321(1)C_{ud,2321}^{(1)}135.7
4CdG,22C_{dG,22}, Cdd,2223C_{dd,2223}, Ced,3312C_{ed,3312}, Cle,1221C_{le,1221}, Clequ,2232(3)C_{lequ,2232}^{(3)}, Cll,1221C_{ll,1221}, Clq,1123(1)C_{lq,1123}^{(1)}, Cqd,1321(8)C_{qd,1321}^{(8)}, Cqu,2333(1)C_{qu,2333}^{(1)}, Cquqd,3321(1)C_{quqd,3321}^{(1)}130.0
5CdW,21C_{dW,21}, CeW,22C_{eW,22}, Ceu,1111C_{eu,1111}, Cledq,3333C_{ledq,3333}, CHq,2213(1)C_{Hq,2213}^{(1)}, CHq,1113(1)C_{Hq,1113}^{(1)}, Cqd,3333(8)C_{qd,3333}^{(8)}, Cquqd,2323(1)C_{quqd,2323}^{(1)}129.5
6CdG,12C_{dG,12}, Clequ,2233(3)C_{lequ,2233}^{(3)}, Clequ,3332(3)C_{lequ,3332}^{(3)}, Clq,2212(1)C_{lq,2212}^{(1)}, Cqd,1213(8)C_{qd,1213}^{(8)}, Cquqd,2323(1)C_{quqd,2323}^{(1)}, Cud,3333(1)C_{ud,3333}^{(1)}127.0
7CdW,21C_{dW,21}, CeB,22C_{eB,22}, Cld,1123(1)C_{ld,1123}^{(1)}, Clequ,3332(3)C_{lequ,3332}^{(3)}, Cll,2332C_{ll,2332}, Clq,1123(1)C_{lq,1123}^{(1)}, Cqd,2312(8)C_{qd,2312}^{(8)}, Cqu,2323(1)C_{qu,2323}^{(1)}125.5
8Cld,3323(1)C_{ld,3323}^{(1)}, Cledq,3333C_{ledq,3333}, Clequ,2223(3)C_{lequ,2223}^{(3)}, Clequ,2233(3)C_{lequ,2233}^{(3)}, Clq,1221(1)C_{lq,1221}^{(1)}, Cqd,2333(8)C_{qd,2333}^{(8)}, Cqd,1213(8)C_{qd,1213}^{(8)}, Cquqd,3121(8)C_{quqd,3121}^{(8)}, Cquqd,3222(8)C_{quqd,3222}^{(8)}124.4
9CdB,12C_{dB,12}, CeW,11C_{eW,11}, Clq,3323(1)C_{lq,3323}^{(1)}, CHl,1123(1)C_{Hl,1123}^{(1)}, Clq,1233(3)C_{lq,1233}^{(3)}, Clq,2333(8)C_{lq,2333}^{(8)}, Cud,1123(1)C_{ud,1123}^{(1)}119.6
10CWC_W, Ced,1123C_{ed,1123}, Clequ,2213(3)C_{lequ,2213}^{(3)}, Clq,1123(1)C_{lq,1123}^{(1)}, CHq,2223(1)C_{Hq,2223}^{(1)}, Cquqd,3121(8)C_{quqd,3121}^{(8)}118.3
11Clequ,2213(3)C_{lequ,2213}^{(3)}, Clequ,2232(3)C_{lequ,2232}^{(3)}, Clq,1112(1)C_{lq,1112}^{(1)}, CHq,1221(1)C_{Hq,1221}^{(1)}, Cqd,2323(1)C_{qd,2323}^{(1)}, Cquqd,3222(1)C_{quqd,3222}^{(1)}, Cquqd,3332(1)C_{quqd,3332}^{(1)}, Cquqd,3112(1)C_{quqd,3112}^{(1)}114.6
12CdG,22C_{dG,22}, Clequ,3332(3)C_{lequ,3332}^{(3)}, Clequ,2233(3)C_{lequ,2233}^{(3)}, Cqd,2333(8)C_{qd,2333}^{(8)}, Clq,1123(1)C_{lq,1123}^{(1)}, Clq,1223(3)C_{lq,1223}^{(3)}113.2
13Cdd,2333C_{dd,2333}, Cld,1112(1)C_{ld,1112}^{(1)}, Clequ,3332(3)C_{lequ,3332}^{(3)}, CHl,1311(1)C_{Hl,1311}^{(1)}, Cqu,2332(1)C_{qu,2332}^{(1)}, Cquqd,1112(8)C_{quqd,1112}^{(8)}111.9
14CdW,22C_{dW,22}, CeB,22C_{eB,22}, Clu,1111C_{lu,1111}, Clq,1321(1)C_{lq,1321}^{(1)}, Cquqd,2323(1)C_{quqd,2323}^{(1)}, Ceu,2223(1)C_{eu,2223}^{(1)}109.7
15Clequ,1133(1)C_{lequ,1133}^{(1)}, Clequ,2212(3)C_{lequ,2212}^{(3)}, Clq,1123(1)C_{lq,1123}^{(1)}, Clq,1213(1)C_{lq,1213}^{(1)}, Cqd,2333(8)C_{qd,2333}^{(8)}, Cqd,2223(3)C_{qd,2223}^{(3)}, Cquqd,3111(1)C_{quqd,3111}^{(1)}, Cud,3312(8)C_{ud,3312}^{(8)}109.2
16CdG,12C_{dG,12}, CdW,31C_{dW,31}, Clequ,2212(3)C_{lequ,2212}^{(3)}, Clequ,1113(3)C_{lequ,1113}^{(3)}, Clq,3312(8)C_{lq,3312}^{(8)}, Clq,3323(8)C_{lq,3323}^{(8)}, Cqd,2323(8)C_{qd,2323}^{(8)}, Cquqd,3312(8)C_{quqd,3312}^{(8)}108.7
17CdB,31C_{dB,31}, CeW,22C_{eW,22}, Clq,1123(1)C_{lq,1123}^{(1)}, Cqd,2323(8)C_{qd,2323}^{(8)}, Cquqd,2321(8)C_{quqd,2321}^{(8)}, Cquqd,2311(8)C_{quqd,2311}^{(8)}107.3
18Cdd,2333C_{dd,2333}, Ced,1122C_{ed,1122}, Clq,1112(1)C_{lq,1112}^{(1)}, Clq,1123(3)C_{lq,1123}^{(3)}, Clq,1113(1)C_{lq,1113}^{(1)}, Cquqd,2131(8)C_{quqd,2131}^{(8)}106.0
19CdB,11C_{dB,11}, CdG,12C_{dG,12}, CdG,23C_{dG,23}, Clequ,2233(1)C_{lequ,2233}^{(1)}, Clq,3312(1)C_{lq,3312}^{(1)}, Cqd,2333(8)C_{qd,2333}^{(8)}, Cquqd,2321(8)C_{quqd,2321}^{(8)}106.0
20CdW,22C_{dW,22}, Cdd,2323C_{dd,2323}, CeW,22C_{eW,22}, Clq,1113(1)C_{lq,1113}^{(1)}, Clq,1123(3)C_{lq,1123}^{(3)}, Clq,1332(3)C_{lq,1332}^{(3)}, Cquqd,3211(8)C_{quqd,3211}^{(8)}, Cquqd,3222(8)C_{quqd,3222}^{(8)}105.7
21Clequ,2233(1)C_{lequ,2233}^{(1)}, Clq,2312(8)C_{lq,2312}^{(8)}, Clq,2323(8)C_{lq,2323}^{(8)}, Clq,1123(3)C_{lq,1123}^{(3)}, Clq,1322(3)C_{lq,1322}^{(3)}, Cquqd,3312(8)C_{quqd,3312}^{(8)}105.0
22CdG,22C_{dG,22}, Clequ,1133(3)C_{lequ,1133}^{(3)}, Clq,1112(8)C_{lq,1112}^{(8)}, Clq,1311(1)C_{lq,1311}^{(1)}, Cqd,2323(8)C_{qd,2323}^{(8)}, Cud,1321(8)C_{ud,1321}^{(8)}104.0
23CHd,1211C_{Hd,1211}, Cqd,2323(8)C_{qd,2323}^{(8)}, Cqd,3312(8)C_{qd,3312}^{(8)}, Cquqd,3112(1)C_{quqd,3112}^{(1)}, Cquqd,3222(8)C_{quqd,3222}^{(8)}103.0
24CeW,11C_{eW,11}, Cqd,1221(1)C_{qd,1221}^{(1)}, Cqd,2332(8)C_{qd,2332}^{(8)}, Cqd,3323(8)C_{qd,3323}^{(8)}, Cquqd,3323(8)C_{quqd,3323}^{(8)}, Cud,1321(8)C_{ud,1321}^{(8)}101.0
25Cdd,2333C_{dd,2333}, Cld,1133C_{ld,1133}, Cledq,3332C_{ledq,3332}, Cqe,1233C_{qe,1233}, Cquqd,1132(8)C_{quqd,1132}^{(8)}, Cquqd,3323(8)C_{quqd,3323}^{(8)}100.5
26Clequ,2222(3)C_{lequ,2222}^{(3)}, Clequ,1111(3)C_{lequ,1111}^{(3)}, Cqd,1112(8)C_{qd,1112}^{(8)}, Clq,1223(1)C_{lq,1223}^{(1)}, Clq,2323(1)C_{lq,2323}^{(1)}, Cquqd,3111(1)C_{quqd,3111}^{(1)}, Cquqd,3323(8)C_{quqd,3323}^{(8)}100.3

Table 2. Distinct SMEFT solution families (J≤0.1J \leq0.1) identified by the RL agent in ten different training runs with different seeds. The solutions are ordered by decreasing Δχ2\Delta\chi^2, and here χSM2=273\chi^2_{\mathrm{SM}} = 273. The 26 solutions contain 122 distinct flavor-specific WCs. We remind that the total number of observables included in the fit are 132. The the definitions of corresponding SMEFT operators are given in Tab. 3.

Bosonic and two-fermion operatorsHiggs–fermion current operators
OW\mathcal{O}_{W}ϵIJKW  νIμW  ρJνW  μKρ\epsilon^{IJK}W^{I\mu}_{\ \ \nu}W^{J\nu}_{\ \ \rho}W^{K\rho}_{\ \ \mu}OHq(1)\mathcal{O}^{(1)}_{Hq}(H†D↔μH)(qˉpγμqr)(H^\dagger\mathrel{\overleftrightarrow{D}}_{\mu}H)(\bar{q}_{p}\gamma^{\mu}q_{r})
OdG\mathcal{O}_{dG}(qˉpσμνTAdr)HGμνA(\bar{q}_{p}\sigma^{\mu\nu}T^{A}d_{r})H G^{A}_{\mu\nu}OHq(3)\mathcal{O}^{(3)}_{Hq}(H†D↔μIH)(qˉpτIγμqr)(H^\dagger\mathrel{\overleftrightarrow{D}}^{I}_{\mu}H)(\bar{q}_{p}\tau^{I}\gamma^{\mu}q_{r})
OdW\mathcal{O}_{dW}(qˉpσμνdr)τIHWμνI(\bar{q}_{p}\sigma^{\mu\nu}d_{r})\tau^{I}H W^{I}_{\mu\nu}OHl(1)\mathcal{O}^{(1)}_{Hl}(H†D↔μH)(lˉpγμlr)(H^\dagger\mathrel{\overleftrightarrow{D}}_{\mu}H)(\bar{l}_{p}\gamma^{\mu}l_{r})
OdB\mathcal{O}_{dB}(qˉpσμνdr)HBμν(\bar{q}_{p}\sigma^{\mu\nu}d_{r})H B_{\mu\nu}OHu\mathcal{O}_{Hu}(H†D↔μH)(uˉpγμur)(H^\dagger\mathrel{\overleftrightarrow{D}}_{\mu}H)(\bar{u}_{p}\gamma^{\mu}u_{r})
OeW\mathcal{O}_{eW}(lˉpσμνer)τIHWμνI(\bar{l}_{p}\sigma^{\mu\nu}e_{r})\tau^{I}H W^{I}_{\mu\nu}OHd\mathcal{O}_{Hd}(H†D↔μH)(dˉpγμdr)(H^\dagger\mathrel{\overleftrightarrow{D}}_{\mu}H)(\bar{d}_{p}\gamma^{\mu}d_{r})
OeB\mathcal{O}_{eB}(lˉpσμνer)HBμν(\bar{l}_{p}\sigma^{\mu\nu}e_{r})H B_{\mu\nu}
(LˉL)(LˉL)(\bar{L}L)(\bar{L}L)(RˉR)(RˉR)(\bar{R}R)(\bar{R}R)
Oll\mathcal{O}_{ll}(lˉpγμlr)(lˉsγμlt)(\bar{l}_{p}\gamma_{\mu}l_{r})(\bar{l}_{s}\gamma^{\mu}l_{t})Odd\mathcal{O}_{dd}(dˉpγμdr)(dˉsγμdt)(\bar{d}_{p}\gamma_{\mu}d_{r})(\bar{d}_{s}\gamma^{\mu}d_{t})
Oqq(1)\mathcal{O}^{(1)}_{qq}(qˉpγμqr)(qˉsγμqt)(\bar{q}_{p}\gamma_{\mu}q_{r})(\bar{q}_{s}\gamma^{\mu}q_{t})Oeu\mathcal{O}_{eu}(eˉpγμer)(uˉsγμut)(\bar{e}_{p}\gamma_{\mu}e_{r})(\bar{u}_{s}\gamma^{\mu}u_{t})
Oqq(3)\mathcal{O}^{(3)}_{qq}(qˉpγμτIqr)(qˉsγμτIqt)(\bar{q}_{p}\gamma_{\mu}\tau^{I}q_{r})(\bar{q}_{s}\gamma^{\mu}\tau^{I}q_{t})Oed\mathcal{O}_{ed}(eˉpγμer)(dˉsγμdt)(\bar{e}_{p}\gamma_{\mu}e_{r})(\bar{d}_{s}\gamma^{\mu}d_{t})
Olq(1)\mathcal{O}^{(1)}_{lq}(lˉpγμlr)(qˉsγμqt)(\bar{l}_{p}\gamma_{\mu}l_{r})(\bar{q}_{s}\gamma^{\mu}q_{t})Oud(1)\mathcal{O}^{(1)}_{ud}(uˉpγμur)(dˉsγμdt)(\bar{u}_{p}\gamma_{\mu}u_{r})(\bar{d}_{s}\gamma^{\mu}d_{t})
Olq(3)\mathcal{O}^{(3)}_{lq}(lˉpγμτIlr)(qˉsγμτIqt)(\bar{l}_{p}\gamma_{\mu}\tau^{I}l_{r})(\bar{q}_{s}\gamma^{\mu}\tau^{I}q_{t})Oud(8)\mathcal{O}^{(8)}_{ud}(uˉpγμTAur)(dˉsγμTAdt)(\bar{u}_{p}\gamma_{\mu}T^{A}u_{r})(\bar{d}_{s}\gamma^{\mu}T^{A}d_{t})
(LˉL)(RˉR)(\bar{L}L)(\bar{R}R)Scalar and tensor structures
Ole\mathcal{O}_{le}(lˉpγμlr)(eˉsγμet)(\bar{l}_{p}\gamma_{\mu}l_{r})(\bar{e}_{s}\gamma^{\mu}e_{t})Oledq\mathcal{O}_{ledq}(lˉpjer)(dˉsqtj)(\bar{l}^{j}_{p}e_{r})(\bar{d}_{s}q^{j}_{t})
Olu\mathcal{O}_{lu}(lˉpγμlr)(uˉsγμut)(\bar{l}_{p}\gamma_{\mu}l_{r})(\bar{u}_{s}\gamma^{\mu}u_{t})Olequ(1)\mathcal{O}^{(1)}_{lequ}(lˉpjer)ϵjk(qˉskut)(\bar{l}^{j}_{p}e_{r})\epsilon_{jk}(\bar{q}^{k}_{s}u_{t})
Old\mathcal{O}_{ld}(lˉpγμlr)(dˉsγμdt)(\bar{l}_{p}\gamma_{\mu}l_{r})(\bar{d}_{s}\gamma^{\mu}d_{t})Olequ(3)\mathcal{O}^{(3)}_{lequ}(lˉpjσμνer)ϵjk(qˉskσμνut)(\bar{l}^{j}_{p}\sigma_{\mu\nu}e_{r})\epsilon_{jk}(\bar{q}^{k}_{s}\sigma^{\mu\nu}u_{t})
Oqe\mathcal{O}_{qe}(qˉpγμqr)(eˉsγμet)(\bar{q}_{p}\gamma_{\mu}q_{r})(\bar{e}_{s}\gamma^{\mu}e_{t})Oquqd(1)\mathcal{O}^{(1)}_{quqd}(qˉpjur)ϵjk(qˉskdt)(\bar{q}^{j}_{p}u_{r})\epsilon_{jk}(\bar{q}^{k}_{s}d_{t})
Oqu(1)\mathcal{O}^{(1)}_{qu}(qˉpγμqr)(uˉsγμut)(\bar{q}_{p}\gamma_{\mu}q_{r})(\bar{u}_{s}\gamma^{\mu}u_{t})Oquqd(8)\mathcal{O}^{(8)}_{quqd}(qˉpjTAur)ϵjk(qˉskTAdt)(\bar{q}^{j}_{p}T^{A}u_{r})\epsilon_{jk}(\bar{q}^{k}_{s}T^{A}d_{t})
Oqu(8)\mathcal{O}^{(8)}_{qu}(qˉpγμTAqr)(uˉsγμTAut)(\bar{q}_{p}\gamma_{\mu}T^{A}q_{r})(\bar{u}_{s}\gamma^{\mu}T^{A}u_{t})
Oqd(1)\mathcal{O}^{(1)}_{qd}(qˉpγμqr)(dˉsγμdt)(\bar{q}_{p}\gamma_{\mu}q_{r})(\bar{d}_{s}\gamma^{\mu}d_{t})
Oqd(8)\mathcal{O}^{(8)}_{qd}(qˉpγμTAqr)(dˉsγμTAdt)(\bar{q}_{p}\gamma_{\mu}T^{A}q_{r})(\bar{d}_{s}\gamma^{\mu}T^{A}d_{t})

Table 3. Warsaw-basis Grzadkowski et al. [ [8] dimension-six operators appearing in the RL solutions. The indices pp, rr, ss, tt denote fermion generations, while jj, kk are SU(2)LSU(2)_{L} indices. The matrices τI\tau^{I} and TAT^{A} denote the SU(2)LSU(2)_{L} and SU(3)cSU(3)_{c} generators, respectively, and ϵjk\epsilon_{jk} is the antisymmetric SU(2)LSU(2)_{L} tensor.

Conclusions and Future Prospects

In order to interpret complex patterns of anomalies with respect to the SM, we have formulated an RL method that can be used to search for viable SMEFT operator hypotheses sequentially. This addresses the central challenge of the combinatorial explosion of possible hypotheses within SMEFT. Our formulation provides a novel approach to this problem by learning an “actor–critic” RL policy.

Importantly, our objective differs from a naive optimization problem in an important respect. We are not interested only in the SMEFT hypothesis that maximizes the improvement in χ2\chi^2. Because several structurally different SMEFT scenarios may provide good descriptions of the same data, mapping this degeneracy is crucial for NP model building in a bottom-up approach. We therefore introduced a diversity reward based on Jaccard similarity, turning the search into a quality–diversity problem in which the RL policy is rewarded both for improving the global fit and for exploring operator sets different from those already discovered.

We tested the framework in a challenging environment containing 132 observables from several sectors, including 24 hypothetical measurements with discrepancies larger than 1.5σ1.5\sigma relative to the SM predictions. The RL agent searched over 912 candidate SMEFT operators and constructed hypotheses sequentially while accounting for the effects of each operator proposal on all observables. The resulting learning behaviour shows that the policy is able to identify 26 high-quality solutions with Jaccard indices below 0.1. Our comparisons with random-subset and random-sequential search baselines provide a direct benchmark of both search efficiency and solution diversity, while independent training seeds offer a systematic way of extending the coverage of the operator space.

Several extensions are natural. The present framework can be applied to larger observable sets and updated experimental measurements, and the quality–diversity objective can be refined to target specific requirements of the physics problem. It will also be useful to investigate how policies trained on one pattern of experimental discrepancies transfer to modified datasets, and whether the learned search strategy can be amortized across related SMEFT inference problems.

More broadly, the approach developed here illustrates how RL can complement conventional phenomenological reasoning by systematically exploring theory spaces whose combinatorial complexity makes exhaustive human-guided analysis increasingly difficult. To facilitate the broader application of our framework, we also provide an open-source Python package, ReinforceSMEFT, that can be readily installed via pip, enabling users to train and apply the RL agent to their own phenomenological analyses.

Acknowledgments

We gratefully acknowledge OpenAI for providing academic access to ChatGPT, which supported aspects of this research. This work was financially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). This research was enabled in part by support provided by the Digital Research Alliance of Canada (https://alliancecan.ca). Computations were performed on the Trillium supercomputer at the SciNet Loken et al. [2010] HPC Consortium. SciNet is funded by Innovation, Science and Economic Development Canada; the Digital Research Alliance of Canada; the Ontario Research Fund: Research Excellence; and the University of Toronto.

Appendix

In Table 3, we provide the definitions of the Warsaw-basis operators that appear in the solutions identified by the RL search. A complete list of Warsaw-basis operators can be found in Refs. Grzadkowski et al. [2010], Aebischer et al. [2018b]. Throughout this work, we adopt the SMEFT down-basis flavor convention. The complete list of operators, including their flavor indices in the WCxf convention, is available on the WCxf Warsaw basis webpage.

References

  1. [2]Jason Aebischer, Andrzej J. Buras, and Jacky Kumar. SMEFT ATLAS: The landscape beyond the Standard Model. Phys. Rept., 1198:1–181, 2026. doi: 10.1016/j.physrep.2026.07.003.DOI
  2. [4]Stephon Alexander, Benjamin Bradley, Loukas Gouskos, and Cooper Niu. Autonomous Discovery of Particle Physics Theories from Experimental Data. 3 2026.arxiv.org/abs/2603.28935
  3. [5]Jason Benjamin Baretz, Max Fieg, Vijay Ganesh, Aishik Ghosh, V. Knapp-Perez, Jake Rudolph, and Daniel Wiesner. Towards AI-assisted neutrino flavor theory design. Commun. Phys., 9(1):227, 2026. doi: 10.1038/s42005-026-02627-2.DOI
  4. [8]B. Grzadkowski, M. Iskrzynski, M. Misiak, and J. Rosiek. Dimension-Six Terms in the Standard Model Lagrangian. JHEP, 10:085, 2010. doi: 10.1007/JHEP10(2010)085.DOI
  5. [9]Ahmed Hammad and Veronica Sanz. Language-Guided Hypotheses Generation for Sparse SMEFT Analyses. 8 2026.arxiv.org/abs/2608.04100
  6. [12]Jacky Kumar, Marianne Bouchard, and David London. Searching for New Physics with Reinforcement Learning. 9 2026.arxiv.org/abs/2609.10382
  7. [15]Brenden K. Petersen, Mikel Landajuela, T. Nathan Mundhenk, Claudio P. Santiago, Soo K. Kim, and Joanne T. Kim. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients, 2021. URL https://arxiv.org/abs/1912.04871.
  8. [16]Shaikh Saad. Large Language Model-Assisted Framework for BSM Model Building. 6 2026.arxiv.org/abs/2606.21316
  9. [17]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402.03300.
  10. [19]Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. MIT Press, 2nd edition, 2018.
  11. [20]Richard S. Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems (NeurIPS), volume 12, pages 1057–1063, 1999.
  12. [21]Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. URL https://arxiv.org/abs/1706.03762.
  13. [23]Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8:229–256, 1992. doi: 10.1007/BF00992696.DOI

Paper details

vlnc:W_5k9at6bs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT

View statement
Looking for related papers…
Contents