SymCircuit: Bayesian Structure Learning for PCs

• SymCircuit is a Bayesian framework that formulates probabilistic circuit generation as sequential decision-making with a learned posterior-inspired policy.
• It replaces greedy top-down search with grammar-constrained Transformer decoding and option-level REINFORCE, greatly boosting sample efficiency.
• The method employs a three-layer uncertainty decomposition—structural, parametric, and leaf—to enable precise and tractable uncertainty estimation.

SymCircuit is a Bayesian structure-learning framework for tractable probabilistic circuits (PCs) that replaces greedy top-down search with a learned generative policy trained by entropy-regularized reinforcement learning. In the formulation introduced in “SymCircuit: Bayesian Structure Inference for Tractable Probabilistic Circuits via Entropy-Regularized Reinforcement Learning” (Ju, 20 Mar 2026), circuit construction is treated as sequential decision-making under a grammar that guarantees valid PCs, the induced optimal policy is a tempered Bayesian posterior over structures, and the resulting system combines posterior-guided structure generation, Transformer-based constrained decoding, and a three-layer uncertainty decomposition. On NLTCS, the reported best run reaches $-6.105$ nats, closing 93% of the gap to LearnSPN, while preliminary results on Plants with 69 variables are presented as evidence of scalability (Ju, 20 Mar 2026).

1. Formal setting and motivation

SymCircuit is motivated by a structural limitation of standard PC learning pipelines: dominant methods such as LearnSPN, ID-SPN, and Strudel use greedy top-down partitioning, make irreversible local decisions, and do not represent uncertainty over alternative structures (Ju, 20 Mar 2026). The framework is designed for PCs whose tractability depends on the usual structural constraints: product nodes combine disjoint scopes, and sum nodes combine children with the same scope. These decomposability and smoothness conditions guarantee exact polynomial-time inference and imply that the circuit output is a multilinear polynomial in the parameters (Ju, 20 Mar 2026).

Within this setting, SymCircuit reframes structure learning as generation of an entire valid circuit rather than recursive commitment to locally chosen splits. The stated objective is not merely to optimize a single structure, but to learn a policy over structures. That shift is consequential because it introduces explicit structural uncertainty and turns search behavior into an object that can itself be amortized across datasets (Ju, 20 Mar 2026).

A recurrent misconception is to treat SymCircuit as a generic neural architecture for PCs. More precisely, it is a structure-learning framework whose central claim is Bayesian: if structure generation is posed as entropy-regularized reinforcement learning over a validity-preserving grammar, then the optimal policy has posterior semantics over the grammar-defined structure space (Ju, 20 Mar 2026).

2. Entropy-regularized reinforcement learning as Bayesian structure inference

The core theoretical construction casts PC generation as an RL-as-inference problem in which actions are grammar tokens and the reward is the empirical average log-likelihood of the data under the generated structure. If $S$ denotes a structure, $R(S)$ its empirical average log-likelihood, $p_0$ a prior policy over structures, and $\alpha>0$ the regularization temperature, then the optimal policy is given by

$$\pi^*(S)=\frac{p_0(S)\exp(R(S)/\alpha)}{Z(\alpha)}, \qquad Z(\alpha)=\sum_{S}p_0(S)\exp(R(S)/\alpha),$$

which can be rewritten as the tempered posterior

$\pi^*(S)\propto p_0(S)\,p(D\mid S)^{1/(N\alpha)}.$

When the temperature is set to

$\alpha=\frac{1}{N},$

the policy recovers the exact Bayesian posterior over structures (Ju, 20 Mar 2026).

This temperature-controlled family interpolates between qualitatively distinct regimes. As $\alpha\to\infty$, the policy approaches the prior; as $\alpha\to 0$, it approaches a MAP point mass; and at $S$0, it coincides with Bayesian model averaging over structures (Ju, 20 Mar 2026). The result is not presented as a heuristic analogy but as an exact characterization in the idealized unconstrained distributional setting.

The paper is explicit, however, that the deployed policy is only an amortized approximation to this target. Its learned policy $S$1 incurs an amortization gap $S$2, and practical training may also involve a temperature mismatch relative to the posterior-identifying regime (Ju, 20 Mar 2026). Accordingly, SymCircuit’s uncertainty estimates should be understood as posterior-inspired rather than as exact full-Bayesian inference in the implemented model.

3. SymFormer and grammar-constrained circuit generation

The policy network used by SymCircuit is SymFormer, a grammar-constrained autoregressive Transformer specialized to PC generation (Ju, 20 Mar 2026). Circuit structures are serialized by a depth-first traversal of a tree, and the grammar is constructed so that any complete generated string is automatically a valid probabilistic circuit. The grammar includes productions of the form

$S$3

with constraints enforcing equal scopes at sum nodes and disjoint scopes at product nodes (Ju, 20 Mar 2026).

Validity is enforced during decoding, not by post hoc repair. At each step the model maintains a grammar state $S$4, computes the valid next-token set $S$5, and masks invalid actions so that

$S$6

As a consequence, every sampled prefix is a valid partial derivation and every completed sample is a valid circuit by construction (Ju, 20 Mar 2026). This sharply distinguishes the method from unconstrained sequence generation approaches that must learn syntactic validity indirectly.

SymFormer also incorporates tree-relative self-attention. Instead of relying only on sequence position, the model augments attention with learned relation-dependent biases for relations in $S$7, together with traversal-aware positional encoding based on tree depth rather than raw token index (Ju, 20 Mar 2026). The purpose is to expose parse-tree structure directly to the policy network while keeping the bias parametrization memory-efficient.

The practical significance of these architectural choices is methodological rather than cosmetic. Grammar-constrained autoregression narrows the support of the policy to the valid PC manifold, while tree-relative attention makes structurally distinct but sequentially nearby contexts separable in the representation. This is the architectural substrate that allows the RL objective to operate on a space of valid tractable circuits rather than on an unrestricted token language (Ju, 20 Mar 2026).

4. Option-level REINFORCE and structural credit assignment

A central training refinement in SymCircuit is option-level REINFORCE. Standard token-level REINFORCE assigns the same scalar reward to every emitted token, but in grammar-constrained PC generation many tokens are effectively forced once earlier structural choices have been made (Ju, 20 Mar 2026). The paper therefore identifies the genuine structural decisions with the sum-node arity choices, indexed by the token positions of type $S$8, and restricts policy-gradient updates to those positions.

The rationale is statistical. If most tokens are deterministic conditional on the grammar state, including them in the score-function estimator adds variance without adding useful credit-assignment signal. Under the paper’s simplifying assumptions, the signal-to-noise ratio of token-level REINFORCE scales like $S$9 in the number of tokens, while the option-level estimator improves SNR by a factor approximately $R(S)$0, where $R(S)$1 is sequence length and $R(S)$2 is the number of structural decisions (Ju, 20 Mar 2026).

On NLTCS, the reported circuits have $R(S)$3 tokens but only $R(S)$4 structural decisions, and the paper reports roughly a $R(S)$5 improvement in gradient SNR (Ju, 20 Mar 2026). The practical effect is a marked increase in sample efficiency: option-level REINFORCE reaches a strong NLTCS model in about 120 circuit evaluations, whereas token-level REINFORCE requires 4,000+ evaluations for weaker performance, summarized as a $R(S)$6 reduction in circuits-to-performance and, in the abstract, as a greater-than-10-times sample-efficiency gain (Ju, 20 Mar 2026).

The paper also reports an option-level actor-critic value head, but notes that on NLTCS its explained variance stayed near zero because there were too few structural decisions for meaningful temporal credit assignment (Ju, 20 Mar 2026). This result narrows the empirical claim: SymCircuit’s main optimization advantage comes from restricting policy-gradient estimation to structurally meaningful actions, not from a strong learned critic.

5. Three-layer uncertainty decomposition

SymCircuit supplements structure learning with a three-layer uncertainty decomposition grounded in the multilinear polynomial structure of PC outputs: structural uncertainty, parametric uncertainty, and leaf uncertainty (Ju, 20 Mar 2026). This component is not incidental; it is the mechanism by which a posterior-like policy over structures is converted into uncertainty-aware prediction.

Structural uncertainty is estimated by model averaging over sampled structures. Given sampled structures $R(S)$7 with fitted parameters, the predictive distribution at a test point is taken as the average over the sampled circuits, and structural variance is estimated by the unbiased sample variance across those predictive values (Ju, 20 Mar 2026). The paper characterizes this as the main epistemic term and notes that, because the policy is tempered and amortized, the estimate is only an approximation to full Bayesian posterior variance.

Parametric uncertainty is derived from the block-diagonal Fisher structure induced by decomposability and smoothness. The Fisher matrix over sum-node parameters decomposes blockwise by sum node, and a delta-method approximation is then used to estimate predictive variance due to parameter uncertainty (Ju, 20 Mar 2026). The asymptotic covariance scales like $R(S)$8, so this term becomes small on large datasets. On NLTCS, with $R(S)$9, the paper reports that parameter uncertainty accounts for only about $p_0$0 of total predictive variance (Ju, 20 Mar 2026).

Leaf uncertainty is propagated exactly for binary leaves via conjugate Dirichlet-Categorical updates. The update uses the top-down flow probability at each leaf, and the paper notes that the necessary sufficient statistics are already computed by Anemone, making the extension essentially free (Ju, 20 Mar 2026). Variance propagation then exploits the PC structure: for product nodes, uncertainty adds exactly in log-space because child scopes are disjoint; for sum nodes, variance is aggregated through the weighted child variances under the paper’s independence assumptions (Ju, 20 Mar 2026).

A common misunderstanding would be to read this decomposition as a generic uncertainty recipe for any probabilistic model. In the presented framework, it depends specifically on the multilinear structure of PCs and on tractability properties that make the relevant covariance approximations and variance propagation analytically manageable (Ju, 20 Mar 2026).

6. Empirical results, scope, and limitations

The strongest empirical results are reported on NLTCS. LearnSPN achieves test log-likelihood $p_0$1 nats, while SymCircuit’s best run reaches $p_0$2 nats (Ju, 20 Mar 2026). The imitation baseline starts at about $p_0$3, and option-level REINFORCE improves it to $p_0$4, which the paper interprets as closing 93% of the gap from the supervised imitation baseline to LearnSPN (Ju, 20 Mar 2026). In a more sample-efficiency-focused comparison, option-level REINFORCE reaches $p_0$5 in 30 epochs with 120 circuit evaluations, while token-level REINFORCE needs around 4,000 evaluations to reach only about $p_0$6 (Ju, 20 Mar 2026).

The Plants results are presented more cautiously. Plants has 69 variables, and after 25 epochs the reported score is $p_0$7 nats, closing 19% of the gap between pretraining and LearnSPN (Ju, 20 Mar 2026). The paper treats these results as preliminary but encouraging evidence that the framework scales beyond the smaller NLTCS setting.

The contribution profile is threefold. First, SymCircuit gives PC structure learning a principled Bayesian interpretation through entropy-regularized RL, with exact posterior recovery at $p_0$8 in the idealized formulation. Second, it provides a concrete generation-and-training stack—SymFormer plus option-level REINFORCE—that guarantees valid circuit generation and improves training efficiency. Third, it adds a tractable decomposition of predictive uncertainty aligned with the multilinear algebra of PCs (Ju, 20 Mar 2026).

The limitations are equally explicit. The learned posterior is only approximate because of amortization and possible temperature mismatch; the grammar restricts expressivity relative to unconstrained greedy methods such as LearnSPN, which may explain the remaining gap on NLTCS; full validation is concentrated on one main benchmark, with Plants still preliminary; and the option-level value head yields little benefit on small circuits (Ju, 20 Mar 2026). A plausible implication is that future work would need to balance stronger posterior fidelity, broader grammar expressivity, and evaluation on a wider set of structured discrete datasets before the method can be regarded as a general replacement for established greedy PC learners.