Knowledge Graphs: The Missing Link in Agentic AI-based Formal Verification
Recent advances in Large Language Models (LLMs) have significantly transformed the landscape of Formal Verification (FV), particularly in the realm of generating SystemVerilog Assertions (SVAs) from natural-language specifications. While the potential for acceleration in FV is promising, challenges persist, primarily due to the ambiguity and incompleteness often inherent in specifications. Furthermore, critical micro-architectural details frequently reside in the Register Transfer Level (RTL), complicating the synthesis process.
Many current methodologies treat specifications and RTL as loosely structured text. This approach diminishes the grounding of specifications to RTL, resulting in semantic mismatches and leading to frequent syntax failures during formal parsing and elaboration. Addressing these limitations requires innovative solutions that enhance the clarity and traceability of the design process.
Introducing Verification-Centric Knowledge Graphs
This article presents a novel approach that leverages a verification-centric Knowledge Graph (KG) to enhance the formal verification process. Constructed from structured Intermediate Representations (IRs) extracted from the specification, RTL, and feedback from formal tools—including syntax diagnostics, Counterexamples (CEXs), and coverage reports—this KG plays a critical role in the synthesis of SVAs.
- Linking Requirements: The KG connects various elements such as requirements, design hierarchy, signals, assumptions, and properties. This linkage provides a traceable and design-grounded context that is essential for effective assertion generation.
- Multi-Agent Workflow: The KG supports a multi-agent workflow that queries and updates its contents to facilitate the generation of SVAs. This workflow drives three key refinement loops:
- Syntax Repair: Guided by tool diagnostics, this loop enhances the syntactic correctness of generated assertions.
- CEX-Guided Correction: Utilizing trace links, this loop focuses on correcting assertions based on counterexamples.
- Coverage-Directed Property Augmentation: This loop aims to enrich the set of properties based on coverage reports.
Evaluation and Results
Extensive evaluations across seven benchmark designs have demonstrated that KG-based context retrieval significantly enhances the grounding of specifications to RTL. The results indicate a consistent production of compilable SVAs, accompanied by low syntax-repair overhead. The findings are promising, with formal coverage levels ranging from 78.5% to 99.4% across different designs.
However, the study also highlights that convergence rates are design-dependent, revealing certain limitations. Complex temporal and arithmetic reasoning continues to present challenges for current LLM capabilities, indicating areas for future research and improvement.
Conclusion
The integration of Knowledge Graphs into the formal verification workflow represents a significant advancement in overcoming longstanding barriers in assertion synthesis. By providing a structured, traceable context for generating SVAs, this approach not only enhances the clarity of specifications but also improves the overall efficiency and effectiveness of the formal verification process. As the field progresses, further exploration into refining these methodologies will be essential in harnessing the full potential of agentic AI in formal verification.
Related AI Insights
- Data Language Models: Revolutionizing Tabular Data AI
- Dynamic Boundary Evaluation: New Benchmark for Language Models
- Boost Non-Thinking Model Performance with Post-Reasoning
- Joint Consistency: Unified Test-Time Aggregation via Energy Minimization
- Debiased Multimodal Personality AI via Dual Causal Intervention
- Controller Class Selection Theory for LLM Action Decisions
- Annotation-Free Logical Consistency Metric for MLLMs
- BioMedArena: Open-Source Toolkit for Biomedical AI Research
- Evaluating Large Language Models for Clinical Action Extraction
- ProCompNav: Navigating Ambiguous Queries with AI
