SpecAlign: Semantic Framework for SystemVerilog Assertion Generation
A novel framework named SpecAlign tackles the issue of semantic alignment in SystemVerilog Assertions (SVAs) produced by Large Language Models (LLMs). Current LLM methodologies prioritize syntactic correctness and formal verification results but often fail to measure the semantic alignment between the generated assertions and their corresponding natural language specifications. This misalignment can undermine confidence and complicate debugging when golden RTL is unavailable. SpecAlign features two iterative alignment loops that evaluate both the properties of natural language and SVAs in relation to the design specification through entailment-based classification. It enhances alignment decisions by creating various reasoning pathways using chain-of-thought prompting and consolidating them via a self-consistency voting system. The analysis of misaligned assertions yields actionable feedback for improvement. This research is available on arXiv under ID 2605.25181.
Key facts
- SpecAlign is a framework for semantic evaluation and refinement of LLM-generated SVAs.
- Existing LLM approaches focus on syntactic validity and formal verification outcomes.
- Semantic alignment between generated assertions and natural language specifications is difficult to quantify.
- Hallucinated or misaligned SVAs can reduce confidence and increase debugging efforts.
- SpecAlign introduces two iterative alignment loops using entailment-based classification.
- It generates multiple reasoning paths using chain-of-thought prompting.
- A self-consistency voting mechanism aggregates reasoning paths.
- Misaligned assertions are analyzed to generate actionable feedback for refinement.
Entities
Institutions
- arXiv