Cost-Driven Synthesis of Sound Abstract Interpreters
By: Qiuhan Gu, Avaljot Singh, Gagandeep Singh
Potential Business Impact:
AI helps check if computer programs are safe.
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.
Similar Papers
Can LLMs Formally Reason as Abstract Interpreters for Program Analysis?
Machine Learning (CS)
Helps computers check computer code for mistakes.
SpeechLLM-as-Judges: Towards General and Interpretable Speech Quality Evaluation
Sound
Helps computers judge how real or good fake voices sound.
Nice to Meet You: Synthesizing Practical MLIR Abstract Transformers
Programming Languages
Makes computer code safer and faster.