What question did this study set out to answer?

The aim is to develop a semantic runtime architecture that improves token efficiency in code generation by separating semantic intent from structural components.

May 16, 2026Open Access

AIR-SR v2: A Semantic Runtime Architecture for Token-Efficient LLM Code Generation

Key Points

The aim is to develop a semantic runtime architecture that improves token efficiency in code generation by separating semantic intent from structural components.
Developed AIR-SR v2 architecture incorporating semantic instruction generation and grammar runtime reconstruction.
Conducted empirical evaluation across 40 benchmark prompts in 9 structured generation domains.
Implemented a three-tier sandbox verification and incremental repair protocol.
Achieved an average token reduction of 81.8%.
Attained a 92.5% validation pass rate across benchmark tests.
Found a strong correlation (r = 0.97) between entropy and compression efficiency.

Abstract

AIR-SR v2 (AI Intermediate Representation — Semantic Runtime) is a semantic runtime architecture for token-efficient large language model (LLM) code generation. The system separates semantic intent from structural boilerplate through entropy decomposition: Hₜotal = Hₛemantic + Hₛyntactic Instead of regenerating full syntax on every request, AIR-SR generates compact semantic instructions (AIR-SR JSON) which are reconstructed deterministically through a grammar-driven runtime across multiple language ecosystems including JavaScript/TypeScript, Python, SQL/Schema, React/JSX, and Go. The architecture consists of: Semantic instruction generation Grammar runtime reconstruction Three-tier sandbox verification Incremental repair protocol Empirical evaluation across 40 benchmark prompts and 9 structured generation domains demonstrates: 81. 8% average token reduction 92. 5% validation pass rate r = 0. 97 entropy–compression correlation AIR-SR operates on the output generation pipeline and is distinct from: hardware compiler IRs (LLVM IR, MLIR-AIR), prompt compression systems, and constrained decoding frameworks. This upload contains the AIR-SR v2 research preprint, benchmark framework results, and system architecture documentation. GitHub Repository: https: //github. com/masoomul786/air-research License: CC BY-NC-SA 4. 0

AIR-SR v2: A Semantic Runtime Architecture for Token-Efficient LLM Code Generation

Key Points

Abstract

Cite This Study