Syntax-Guided Provider/Decider Architecture

• Syntax-Guided Provider/Decider Architecture is a structured search framework that decouples candidate generation (provider) from semantic decision-making (decider) to ensure syntactic validity and correctness.
• It leverages multiple provider mechanisms like rule predictors, tree copiers, and subtree locators alongside a mixture-of-experts decider to integrate scores and refine candidate solutions.
• The approach increases efficiency in automated program repair and synthesis by reducing search space, guaranteeing minimality, completeness, and accelerating performance through learned grammar filtering.

The syntax-guided provider/decider architecture is a principled approach to structured search, learning, and patch generation, in which the roles of candidate-generation (provider) and candidate-selection or refinement (decider) are explicitly separated and often tightly integrated. Originating in syntax-guided synthesis (SyGuS) and neural program repair, the paradigm enables efficient, correct-by-construction solutions to problems where syntactic constraints and semantic requirements intersect. This architecture underlies neural edit decoders for automated program repair (Zhu et al., 2021), concolic synthesis frameworks in SyGuS (Huang et al., 2018), and neural-augmented symbolic search for grammar filtering (Morton et al., 2020).

1. Conceptual Foundations

In its canonical formulation, the provider/decider split arises naturally in problems that combine (i) a combinatorial or continuous space of possible solutions, governed by a domain-specific syntax, and (ii) a semantic or correctness specification. The provider component is responsible for proposing candidate subspaces—such as partial derivations, production sets, or structural bounds—admissible under the syntactic constraints. The decider then evaluates, selects among, scores, or refines these candidates according to semantic adequacy (e.g., satisfying a logical specification, maximizing a neural score, or passing test cases).

In syntax-guided program repair (e.g., Recoder (Zhu et al., 2021)), the provider emits expansions for edit scripts according to a context-free grammar (CFG) of possible edits, along with operator- and subtree-level copying mechanisms. In concolic SyGuS synthesizers (Huang et al., 2018), the provider proposes structure bounds (e.g., decision-tree height), and the decider executes a symbolic (CEGIS) procedure to decide feasibility. In grammar filtering (Morton et al., 2020), a neural network provider filters unpromising productions, and a decider runs a symbolic solver on the pruned grammar.

2. Provider Components and Mechanisms

Providers enumerate or sample from a space of syntactically well-formed candidates, often exploiting context or local structure. In Recoder (Zhu et al., 2021), three distinct providers operate:

• Rule Predictor: Applied to non-terminals such as $\mathit{Edits}$, $\mathit{Insert}$, and various identifier types. For a decoder state $d \in \mathbb{R}^D$, it computes unnormalized scores for each grammar production:

$$s^r_j = w_j^{\mathsf{T}} d + b_j \,,\quad j=1, \dots, N_r\,,$$

followed by masked softmax:

$$P_r(j \mid d) = \frac{\exp(s^r_j)}{ \sum_k \exp(s^r_k) }\,,$$

where the mask restricts choices to productions matching the current non-terminal.

• Tree Copier: For host-language non-terminals, all subtrees in the context whose root matches the NT type are enumerated. Scores use a pointer-network mechanism:

$$\theta^t_m = v^{\mathsf{T}} \tanh( W_1 d + W_2\, \mathrm{ctx}(t_m) )$$

with softmax normalization over candidate subtrees.

• Subtree Locator: For $\mathit{Modify}$, subtrees of the faulty statement exceeding size $1$ are scored as in the tree copier.

Each provider conditions on the full decoder state and is responsible for a partition of the grammar's non-terminals. In neural grammar filtering (Morton et al., 2020), the provider is a neural multi-label classifier, predicting “noncriticality” of SyGuS grammar productions from input-output example embeddings and historical timing data.

In enumerative/symbolic SyGuS (Huang et al., 2018), the provider is a (possibly learning-augmented) process that proposes shape bounds (e.g., increasing decision tree height $h$).

3. Decider Components and Integration

The decider integrates outputs from (possibly multiple) eligible providers at each step, using mixture-of-experts or selection policies. In Recoder (Zhu et al., 2021), a learned mixture over providers is defined for each decoder state $d$: $\mathit{Insert}$0 with logic masking to zero out inapplicable providers. Expansion probabilities for all candidates $\mathit{Insert}$1 are given by: $\mathit{Insert}$2 The decider thus arbitrates at each non-terminal, supporting context-sensitive, type-directed, and structure-preserving generation. In symbolic SyGuS (Huang et al., 2018), the decider is a CEGIS loop at fixed candidate structure, accepting or rejecting complete solutions according to specification satisfaction.

In grammar filtering (Morton et al., 2020), the decider is a standard CEGIS-based SyGuS solver, invoked either on the filtered grammar or, via fallback, on the full grammar, ensuring soundness and empirical completeness.

4. Decoding Algorithms and Workflow

The decoding procedure in these architectures is typically a prioritized search or beam search over partial derivations. For instance, Recoder (Zhu et al., 2021) employs beam search over partial edit scripts: $\mathit{Insert}$6 In symbolic SyGuS, the top-level loop enumerates shape parameters (e.g., decision-tree heights $\mathit{Insert}$3), each of which triggers a local CEGIS decision procedure at fixed structure (Huang et al., 2018).

Provider/pruning-based PBE synthesis (Morton et al., 2020) pre-filters the grammar, then invokes the standard symbolic decider, with a fallback to the original grammar if no solution emerges within a sub-timeout.

5. Theoretical Guarantees and Empirical Outcomes

The separation of candidate space exploration (provider) and correctness or semantic search (decider) offers several key advantages:

• Syntactic soundness: Expansions produced by the mutually-constrained providers and mixture-based decider obey a CFG, guaranteeing well-formedness (e.g., in Recoder (Zhu et al., 2021), all edit scripts yield syntactically valid programs).
• Compact search space: Small program repairs or simple grammar prunings collapse search spaces dramatically. Recoder can express single-token changes as single-edit commands rather than lengthy token emissions, making the search more tractable.
• Identifier generalization and placeholders: Recoder’s grammar is extended to emit $\mathit{Insert}$4, supporting project-specific identifier instantiation without vocabulary explosion.
• Minimality and completeness: In height-guided synthesis (Huang et al., 2018), the provider’s structure enumeration guarantees that minimal (by height) solutions will be found if they exist.
• Efficiency: Grammar filtering through provider/decider accelerates CEGIS-based synthesis by up to 47.65% in aggregate runtime on SyGuS PBE benchmarks (Morton et al., 2020), with virtually no loss in completeness.

Empirical results across benchmarks such as Defects4J (Recoder vs. TBar, SimFix), IntroClassJava, and QuixBugs, as well as SyGuS PBE-Strings competitons, demonstrate state-of-the-art or superior performance for provider/decider-based systems in both neural and symbolic domains.

6. Extensions, Fragment Characterizations, and Limitations

Specialized fragments of synthesis problems admit purely symbolic, complete providers/deciders. In SyGuS:

• Strong Single Invocation (SSI): Problems where the target function $\mathit{Insert}$5 is only ever invoked once per formula atom, admitting a polynomial-time provider/decider pipeline (via quantifier elimination) (Huang et al., 2018).
• Acyclic Translational (AT) Invariant Synthesis: For loop invariants under guarded affine translation, the associated transition graph is acyclic, and the invariant is synthesized in polynomial time (Huang et al., 2018).

Possible extensions include grammar-guided or size-bounded enumeration, learning-based providers, and parallelization of provider tasks across the candidate space. Notable limitations arise from exponential search in structure-bound enumeration, and coverage is linked to the expressivity of the underlying grammar and learning capacity of the provider.

7. Comparative Summary and Research Impact

The syntax-guided provider/decider paradigm unifies disparate search methodologies by explicitly decoupling search space construction and semantic decision layers. The architecture has enabled:

• First demonstration of a deep learning-based APR system (Recoder) outperforming traditional approaches on Defects4J (Zhu et al., 2021).
• Efficient, provably complete synthesis on entire classes of linear arithmetic and loop invariant problems by symbolic provider/decider composition (Huang et al., 2018).
• A nearly twofold acceleration of SyGuS PBE solvers through learned grammar filtering (Morton et al., 2020).

This structural framework continues to inform both theoretical advancements in automated reasoning and practical improvements in neural program synthesis and repair. The explicit provider/decider boundary supports modular enhancements, integration of learned and symbolic components, and analysis of workflow bottlenecks and guarantees across the synthesis and repair landscape.