acceptodds
Desk rejected at ICLR 2027

Executable Semantic Alignment for Specification-Driven Program Generation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language models (LLMs) have substantially improved natural- language-to-program generation, enabling specification-driven program synthesis. However, specification-driven program generation still faces the challenge of preserving execution semantics defined by specifications. In this work, we study the semantic-fidelity problem in specification- to-program transformation and propose Executable Semantic Alignment (ESA). ESA introduces explicit execution-semantic representations, recov- ers atomic execution behaviors through parameterized semantic interfaces, and decomposes compositional semantic recovery into parsing processes for different semantic blocks to construct Logical IR. We construct a cross-domain dataset containing 50 standard documents and 3,055 specification clauses. We find that although natural-language specifications exhibit high linguistic diversity, their corresponding execu- tion behaviors show substantial semantic convergence. On this dataset, we evaluate semantic fidelity from three perspectives: end-to-end program generation, atomic execution-semantic recovery, and compositional seman- tic reconstruction. Experimental results show that ESA effectively improves semantic fidelity in specification-driven program generation and validates the effectiveness of explicit semantic representations and structured Logical IR construction.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Large Language Models, Specification-Driven Program Generation, Semantic Alignment, Program Synthesis

Related papers

Discussion (0)

Sign in to comment.