Knowledge Graphs Enhance Agentic AI for Formal Verification
A new arXiv paper (2605.06434) proposes using knowledge graphs (KGs) to improve agentic AI-based formal verification. Large language models (LLMs) can generate SystemVerilog Assertions (SVAs) from natural-language specs, but face ambiguity, incomplete details, and syntax errors. The KG, built from structured intermediate representations (IRs) of specs, RTL, and formal-tool feedback (syntax diagnostics, counterexamples, coverage reports), links specification-to-RTL grounding, reducing semantic mismatches and failures. This work addresses limitations of treating specs and RTL as loosely structured text.
Key facts
- arXiv paper 2605.06434 proposes knowledge graphs for agentic AI formal verification.
- LLMs generate SystemVerilog Assertions from natural-language specifications.
- Specifications are often ambiguous or incomplete.
- Critical micro-architectural details reside in Register Transfer Level (RTL).
- Existing approaches treat specification and RTL as loosely structured text.
- Knowledge graph is constructed from structured intermediate representations.
- IRs extracted from specification, RTL, and formal-tool feedback.
- Feedback includes syntax diagnostics, counterexamples, and coverage reports.
Entities
Institutions
- arXiv