grammar engineering

To address the rugged fitness landscape and the lack of semantic guidance in traditional syntactic operators for test-driven program synthesis, this paper proposes a semantics-aware genetic programming framework. Methodologically, it (1) employs graph neural networks to model instruction-data interactions within program abstract syntax graphs, generating node saliency maps that quantify semantic importance; (2) introduces the notion of “latent value” to prioritize low-fitness yet evolutionarily promising individuals during search; and (3) designs semantics-sensitive mutation and crossover operators. Evaluated on multiple symbolic regression benchmarks, the approach achieves up to 42% faster convergence compared to classic tree-based genetic programming, while significantly improving solution quality stability. Ablation studies confirm the distinct and measurable contributions of each component.

This work addresses the limitations of existing program evolution methods, which struggle to effectively control the magnitude of behavioral variation and rely on syntactic edit distances that poorly capture semantic change. To overcome these challenges, the authors propose a large language model–based program evolution framework that introduces natural language as a guided, executable intermediate representation. By fine-tuning Qwen3-8B with task-conditioned supervision and orthogonal Direct Preference Optimization (oDPO), the framework enables conditional semantic variation and bidirectional compilation between natural language and a domain-specific language (GPTL). Evaluated under a fixed budget of 252 search iterations, the approach significantly improves behavioral alignment and search efficiency, with natural language representations achieving optimal fitness retention and enabling synergistic optimization of behavioral displacement and semantic fidelity.

Existing structured-input processing systems often lack complete and up-to-date syntactic and semantic specifications; while syntax mining has focused primarily on parsing structure, semantic recovery remains unaddressed. Method: We propose the first approach to automatically infer attribute grammars from recursive-descent parser implementations. Our method combines dynamic execution tracing and program instrumentation to capture runtime behavior, augmented by control-flow analysis and grammar-driven semantic mapping, thereby precisely associating parsing operations with productions and extracting semantic actions. Contribution/Results: This work pioneers syntax mining at the semantic level, enabling fully automated generation of executable attribute grammars that faithfully model input-processing logic. Evaluation across multiple real-world programs demonstrates that the inferred grammars accurately reproduce original parser behavior—enabling novel applications in reverse engineering, specification documentation, and security analysis.

This work addresses the performance bottleneck in search-based program synthesis caused by the high computational cost of fine-grained abstract semantics, which, while effective at pruning incorrect programs, hinders overall efficiency. To overcome this limitation, the authors propose an offline pre-synthesis approach that first constructs a tree automaton over the input space to precisely capture the abstract semantics of the domain-specific language (DSL). This automaton enables the generation of an efficient pruning oracle that decouples abstract semantic reasoning from the online synthesis process. By doing so, the method achieves, for the first time, highly efficient synthesis under fine-grained abstract semantics. Empirical evaluations demonstrate substantial performance improvements over state-of-the-art techniques across three diverse domains: SQL query synthesis, string transformation, and matrix manipulation.

Formal program specifications are notoriously difficult, error-prone, and inefficient to write manually. To address this, we propose a two-stage LLM-driven approach: dialogue-guided specification synthesis followed by mutation-based verification. First, multi-turn dialogues model complex semantic requirements; second, four mutation operators—insertion, replacement, deletion, and reordering—enable verifiability-driven selection, eliminating reliance on rigid templates or syntactic grammars. Our method integrates code understanding, prompt engineering, and heuristic verifiability assessment. Evaluated on SV-COMP and a custom Java benchmark comprising 385 programs, it generates 279 verifiable specifications. These achieve significantly higher completeness and accuracy than pure-LLM baselines and classical tools (e.g., Houdini, Daikon). To our knowledge, this is the first approach to achieve both high coverage and formal verifiability in fully automated specification generation.

This work addresses key limitations in existing large language model (LLM)-based evolutionary search, which struggles to distinguish semantically equivalent yet syntactically diverse programs and fails to effectively preserve strategic potential or detect saturation within strategy families. To overcome these challenges, the authors propose introducing a strategy-space layer into program evolution, elevating natural language strategy descriptions to first-class evolutionary states. By integrating strategy representation, hierarchical experience retrieval, and strategy-landscape navigation, the method enables explicit organization, memory, and exploration at the strategy level. Combining LLMs with evolutionary algorithms, strategy clustering, behavior-complementarity-based retrieval, and summarization mechanisms, the approach significantly outperforms baseline methods across tasks in mathematical algorithm discovery, system optimization, and agent framework design, achieving a relative improvement of 21% in open-ended system optimization.

This work addresses the problem of syntactic ambiguity in context-free grammars, where ambiguities can lead to unintended scoping, operator precedence, and associativity that deviate from language design intent. To resolve this, the authors propose an example-based disambiguation method that synthesizes user-preferred disambiguation rules from programming examples and the original grammar using a novel tree automaton learning algorithm. They further introduce an efficient tree automaton intersection algorithm that significantly compresses the resulting specification, ensuring both readability and compatibility with mainstream parser generators. The implemented tool, Greta, successfully eliminates ambiguities across multiple case studies, producing unambiguous, canonical grammars suitable for standard parser generators, while the approach is theoretically guaranteed to be correct.

This work addresses the limitations of existing large language model (LLM)-driven heuristic design methods in combinatorial optimization, which often rely on manual trial-and-error or domain-specific knowledge and lack a systematic mechanism for improvement. To overcome this, the authors propose a structured framework that formalizes heuristic discovery as a language-guided program optimization process, comprising three modular phases: forward evaluation, backward feedback, and program update. This design enables an iterative and composable optimization workflow, unifying and generalizing prior approaches while allowing flexible enhancements through modularity. Empirical evaluation across four real-world combinatorial optimization tasks demonstrates that the proposed method significantly outperforms baseline techniques, achieving up to a 0.17 improvement in the QYI metric on unseen test instances.

This work addresses the poor locality and low sample efficiency in genetic programming, where minor syntactic mutations often induce drastic behavioral changes. To mitigate this, the authors construct a continuous program space endowed with behavioral meaning and introduce a semantics-aware mutation operator. This operator establishes a behaviorally trustworthy neighborhood through block-decomposition embeddings to quantify locality, and combines flow-model-guided directional sampling with geometric compilation-based mutations within semantically aligned subspaces. Evaluated under identical evolutionary strategies and evaluation budgets, the proposed method discovers superior trading strategies using only one-tenth of the evaluations required by baseline approaches, significantly improving the median out-of-sample Sharpe ratio.

This work addresses key limitations in existing large language model (LLM)-based automated program repair methods, particularly their shortcomings in repair diversity, semantic relationship modeling, partial repair composition, and utilization of failure feedback. To overcome these challenges, the paper proposes EvolRepair, a novel framework that introduces a semantic evolution mechanism into LLM-driven program repair for the first time. EvolRepair leverages behaviorally consistent repair populations, semantic-aware genetic operators, and structured execution feedback to enable family-based repair inference, cross-individual complementary synthesis, and dynamic search redirection guided by failure patterns. Experimental results demonstrate that EvolRepair significantly outperforms current LLM-based repair approaches in both repair success rate and solution quality.