HAVEN

Tool

First seen 7/12/2026

Last seen 7/12/2026

Evidence 5 chunks

NEIGHBORHOOD

No graph connections found for this entity yet. It may appear in future ingestion runs.

explore full graph →

RELATIONSHIPS

26 connections

LLM Agent uses → 99% 3e

HAVEN leverages LLM agents to analyze design specifications and coverage gap reports.

Rule-Based Code Generation uses → 99% 3e

HAVEN delegates all code generation to a rule-based CodeGen that guarantees SystemVerilog syntactic correctness.

AutoBench ← compares with 90% 2e

HAVEN is compared against AutoBench as a prior LLM-based testbench generation approach.

CorrectBench ← compares with 90% 2e

HAVEN is compared against CorrectBench as a prior LLM-based testbench generation approach.

ConfiBench ← compares with 90% 2e

HAVEN is compared against ConfiBench as a prior LLM-based testbench generation approach.

Coverage Gap Report uses → 98% 2e

HAVEN iteratively uses LLM agents to analyze coverage gap reports to generate targeted DSL sequences.

Functional Coverage evaluates → 98% 2e

HAVEN is evaluated on functional coverage, achieving 87.9% on average.

Code Coverage evaluates → 98% 2e

HAVEN is evaluated on code coverage, achieving 90.6% on average.

Wishbone Protocol evaluates → 93% 2e

HAVEN is evaluated on IP designs using Wishbone interface protocol.

Protocol-Aware Sequence Domain-Specific Language introduces → 99% 2e

HAVEN introduces a Protocol-Aware Sequence DSL that decomposes sequences into fine-grained step types.

Integrated Circuit Verification evaluates → 95% 2e

HAVEN is evaluated in the context of IC verification, achieving state-of-the-art coverage results.

Universal Verification Methodology implements → 98% 2e

HAVEN generates UVM testbenches and sequences following the Universal Verification Methodology framework.

Structured JSON Artifact uses → 97% 2e

HAVEN uses LLM agents to extract information into structured JSON artifacts (UVM Blueprint and Protocol Flows).

Jinja2 Template Engine uses → 99% 2e

HAVEN uses pre-defined, protocol-specific Jinja2 templates to generate all UVM components.

Large Language Model uses → 98% 2e

LLMs are used by HAVEN for structured information extraction but not for direct HDL code generation.

Direct Interface Protocol evaluates → 93% 1e

HAVEN is evaluated on IP designs using Direct interface protocol.

UVM Blueprint uses → 97% 1e

HAVEN's LLM agents produce a UVM Blueprint as one of the structured JSON artifacts.

Protocol Flows uses → 97% 1e

HAVEN's LLM agents produce Protocol Flows as one of the structured JSON artifacts.

Compile-Fix Loop uses → 96% 1e

HAVEN employs a compile-fix loop to ensure the testbench compiles correctly before Stage 2.

Bus Functional Model uses → 95% 1e

HAVEN generates peripheral Bus Functional Models as part of the UVM testbench components.

Constrained Random Stimulus uses → 93% 1e

HAVEN uses predefined constrained random DSL patterns to achieve coverage without LLM involvement.

Field Value Sweep uses → 93% 1e

HAVEN uses field value sweep patterns as predefined DSL sequences.

Toggle Pattern uses → 93% 1e

HAVEN uses toggle patterns as predefined DSL sequences.

FIFO Stress Test uses → 93% 1e

HAVEN uses FIFO stress tests as predefined DSL sequences.

UVM2 compares with → 90% 1e

HAVEN is compared against UVM2 as a prior template-assisted LLM testbench generation approach.

AXI4-Lite Protocol evaluates → 93% 1e

HAVEN is evaluated on IP designs using AXI4-Lite interface protocol.