Cost-Driven Synthesis of Sound Abstract Interpreters
Gu, Qiuhan, Singh, Avaljot, Singh, Gagandeep
–arXiv.org Artificial Intelligence
Constructing abstract interpreters that provide global soundness guarantees remains a major obstacle in abstract interpretation. We investigate whether modern LLMs can reduce this burden by leveraging them to synthesize sound, non-trivial abstract interpreters across multiple abstract domains in the setting of neural network verification. We formulate synthesis as a constrained optimization problem and introduce a novel mathematically grounded cost function for measuring unsoundness under strict syntactic and semantic constraints. Based on this formulation, we develop a unified framework that unifies LLM-based generation with syntactic and semantic validation and a quantitative cost-guided feedback mechanism. Empirical results demonstrate that our framework not only matches the quality of handcrafted transformers, but more importantly, discovers sound, high-precision transformers for complex nonlinear operators that are absent from existing literature.
arXiv.org Artificial Intelligence
Nov-18-2025
- Country:
- Asia
- Middle East
- Israel > Haifa District
- Haifa (0.04)
- Jordan (0.04)
- Israel > Haifa District
- Singapore > Central Region
- Singapore (0.04)
- Middle East
- North America
- Asia
- Genre:
- Research Report > New Finding (0.66)
- Industry:
- Information Technology (0.45)
- Technology: