GoGoTB: Agentic RTL Verification with Specification-Grounded Coverage Closure GoGoTB：基于规范接地覆盖关闭的智能RTL验证

TL;DR

• GoGoTB is an agentic framework for automated RTL verification that uses LLMs with shared context across components to avoid interface mismatches and disconnected coverage.
• It features three subsystems: an execution control layer separating deterministic enforcement from LLM reasoning, an evolvable knowledge system for on-demand expertise dispatch, and a specification-grounded coverage closure framework anchoring bins to named behaviors.
• Tested on 8 RTL designs without human intervention, it achieved 100% environment generation success and high coverage metrics (98.4% line, 97.2% branch, 97.0% toggle, 83.2% functional), outperforming prior work.
• The approach enables end-to-end verification closure by ensuring each residual coverage gap has a diagnosable root cause and targeted remedy tied to specifications.
• This represents a significant step toward fully automated IC verification using agentic AI systems with contextual coherence and specification alignment.

Why It Matters

This work addresses critical pain points in semiconductor design automation where manual verification consumes excessive resources and missed bugs lead to costly silicon respins. By demonstrating that agentic frameworks can achieve near-complete coverage without human intervention while maintaining specification traceability, it establishes a new benchmark for reliability in AI-driven EDA tools. For practitioners, this signals a shift from isolated LLM applications to coordinated multi-agent systems capable of handling complex engineering workflows with contextual awareness.

Technical Details

• Agentic Execution Control Layer: Separates deterministic rule enforcement from LLM reasoning at every tool boundary and stage transition, ensuring compliance with verification protocols while allowing flexible problem-solving via language models.
• Evolvable Knowledge System: Dynamically dispatches methodology-specific and design-specific expertise as needed during verification tasks, enabling adaptive responses to diverse RTL challenges without static programming.
• Specification-Grounded Coverage Closure Framework: Anchors each coverage bin to explicitly named specification behaviors, transforming abstract coverage gaps into actionable items with identifiable root causes and remediation strategies.
• End-to-End Verification Pipeline: Integrates testbench generation, stimulus creation, simulation, and coverage analysis within a single coherent workflow, eliminating siloed component generation common in prior LLM approaches.
• Benchmark Results: Achieved 100% successful environment generation across 8 RTL designs with average coverage of 98.4% line, 97.2% branch, 97.0% toggle, and 83.2% functional—metrics previously unattainable by non-agentic methods on these benchmarks.

Industry Insight

The success of GoGoTB suggests that agentic architectures will become foundational for next-generation EDA automation, particularly in domains requiring long-horizon planning and cross-component coordination like chip verification. Companies investing in LLM-based verification should prioritize building shared-context frameworks rather than point solutions to avoid integration pitfalls and ensure specification traceability. As coverage targets rise toward 95%+ functional levels, the ability to diagnose and close gaps systematically—as demonstrated here—will differentiate viable production-ready tools from experimental prototypes, accelerating adoption cycles in semiconductor manufacturing.

TL;DR

• GoGoTB 提出了一种基于智能体的 RTL 验证框架，通过三个子系统实现端到端验证闭合。
• 执行控制层将确定性执行与 LLM 推理分离，确保工具边界上的严格约束。
• 知识系统按需调度方法和设计特定 expertise，提升上下文一致性。
• 覆盖框架将每个覆盖 bin 锚定到命名规范行为，使未覆盖项可诊断、可修复。
• 在 8 个 RTL 设计上无需人工干预，实现 100% 环境生成成功率及高覆盖率（线 98.4%、分支 97.2%、翻转 97.0%、功能 83.2%）。

为什么值得看

该工作解决了现有 LLM 驱动验证中上下文断裂、接口不匹配和覆盖与规范脱节的核心痛点，首次在无干预下完成完整验证环境构建并达成显著覆盖率，为 AI 赋能硬件验证提供了可落地的范式。其“规范锚定覆盖”机制尤其具有行业革新潜力，可直接降低芯片返工风险与工程成本。

技术解析

• 架构设计：GoGoTB 由三大子系统构成——智能体执行控制层、进化式知识系统、规范接地覆盖闭合模块，形成闭环验证流程。
• 执行控制层：在每个工具和阶段边界强制分离确定性逻辑（如语法检查、格式校验）与 LLM 生成的创造性决策，避免错误扩散。
• 知识系统：动态检索并注入方法论规则与设计上下文，支持多轮对话中的状态记忆与跨组件协调，解决单轮调用导致的接口不一致问题。
• 覆盖闭合机制：所有覆盖率 bin 均绑定具体规范行为描述，任何未覆盖项均可追溯至缺失的功能点，并自动生成针对性测试用例或修正建议。
• 实验结果：在 8 个独立 RTL 基准上零人工介入，环境生成成功率 100%，平均功能覆盖率达 83.2%，显著优于先前同类方法。

行业启示

• 推动硬件验证从“辅助辅助”向“自主代理”演进，未来可集成进 EDA 流水线实现全自动签核前验证。
• “规范接地”模式将成为下一代 AI 验证工具的标准配置，确保输出结果可解释、可审计、符合形式化要求。
• 建议芯片厂商优先在中小型 IP 模块试点此类框架，逐步积累训练数据与领域知识库，降低迁移至全自动化验证的风险。