00

GRPO with State Mutations: Improving LLM-Based Hardware Test Plan Generation

Paper ☆ WATCH ⎘ SHARE

First seen 9/3/2026

Last seen 9/3/2026

Evidence 13 chunks

01

NEIGHBORHOOD

37 nodes · 53 edges

graph · GRPO with State Mutations: Improving LLM-Based Hardware Test Plan Generation · depth=1

02

RELATIONSHIPS

36 connections

ScaleRTL dataset uses → 100% 3e evidence

The paper uses the ScaleRTL dataset as the source of RTL codes for evaluation and training.

Decomposed prompting: a modular approach for solving complex tasks mentions → 100% 2e evidence

The paper cites Decomposed Prompting to motivate its two-stage task decomposition.

two-stage test plan generation framework introduces → 100% 2e evidence

The paper introduces a two-stage framework for test plan generation and testbench execution.

GRPO with State Mutation introduces → 100% 2e evidence

The paper proposes GRPO-SMu as a novel RL training method.

tree-based branching mutation strategy introduces → 100% 2e evidence

The paper presents a novel tree-based branching mutation strategy for training data curation.

supervised fine-tuning uses → 100% 2e evidence

The paper uses supervised fine-tuning as part of its training methodology.

reinforcement learning for language models uses → 100% 2e evidence

The paper applies reinforcement learning to teach LLMs from verification outcomes.

Mutation Testing uses → 100% 2e evidence

The paper uses mutation testing to generate buggy RTL variants for evaluation and training.

Random Test Generation uses → 100% 2e evidence

The paper uses a random test generator to validate mutations and evaluate testbenches.

reward modeling uses → 100% 2e evidence

The paper uses a nuanced reward model operating on a 0-3 scale to distinguish test plan quality.

DeepSeek-R1 evaluates → 100% 2e evidence

The paper benchmarks DeepSeek-R1 on RTL verification stimuli generation.

Claude-4.0-Sonnet evaluates → 100% 2e evidence

The paper benchmarks Claude-4.0-Sonnet on RTL verification stimuli generation.

DeepSeek-R1-distill-Qwen-7B evaluates → 100% 2e evidence

The paper uses and evaluates DeepSeek-R1-distill-Qwen-7B as its base small language model.

LLaMA-3.1-405B uses → 100% 2e evidence

The paper uses LLaMA-3.1-405B for stage 2 testbench generation and SFT data curation.

ScaleRTL: scaling LLMs with reasoning data and test-time compute for accurate RTL code generation mentions → 100% 2e evidence

The paper cites and uses the ScaleRTL paper's dataset.

BugGen: a self-correcting multi-agent LLM pipeline for realistic RTL bug synthesis mentions → 100% 2e evidence

The paper cites BugGen for its approach to RTL bug synthesis and compares against it.

Comprehensive Verilog Design Problems (CVDP) benchmark mentions → 100% 2e evidence

The paper uses CVDP benchmark for deriving fine-grained mutation operators.

DAPO: an open-source LLM reinforcement learning system at scale mentions → 100% 2e evidence

The paper cites DAPO for its token-level loss approach and adapts it for sparse reward handling.

Dimple Vijay Kochar authored by → 100% 1e evidence

The paper lists Dimple Vijay Kochar as a co-author affiliated with MIT.

ScaleRTL-32B evaluates → 100% 1e evidence

The paper evaluates ScaleRTL-32B as an RTL-specific fine-tuned baseline.

Haoxing Ren authored by → 100% 1e evidence

The paper lists Haoxing Ren as a co-author affiliated with NVIDIA Research.

Brucek Khailany authored by → 100% 1e evidence

The paper lists Brucek Khailany as a co-author affiliated with NVIDIA Research.

Chenhui Deng authored by → 100% 1e evidence

The paper lists Chenhui Deng as a co-author affiliated with NVIDIA Research.

MIT authored by → 95% 1e evidence

The paper's first author is affiliated with MIT.

NVIDIA Research authored by → 100% 1e evidence

Multiple co-authors of the paper are affiliated with NVIDIA Research.

DeepSeekMath: pushing the limits of mathematical reasoning in open language models mentions → 100% 1e evidence

The paper cites DeepSeekMath as the origin of the GRPO technique.

Claude3.7-Sonnet uses → 100% 1e evidence

The paper uses Claude3.7-Sonnet for SFT data curation and reasoning trace generation.

Nathaniel Pinckney authored by → 100% 1e evidence

The paper lists Nathaniel Pinckney as a co-author affiliated with NVIDIA Research.

testbench template uses → 100% 1e evidence

The paper uses a testbench template as a skeleton stub in stage 2 to reduce syntax errors.

SFT training dataset uses → 100% 1e evidence

The paper trains DeepSeek-R1-distill-Qwen-7B on the curated SFT training dataset.

GRPO-SMu training dataset uses → 100% 1e evidence

The paper uses the GRPO-SMu training dataset constructed via tree-based mutation strategy.

functional equivalence checking uses → 100% 1e evidence

The paper uses functional equivalence checking to validate mutations.

VeriContaminated: assessing LLM-driven Verilog coding for data contamination mentions → 90% 1e evidence

The paper cites VeriContaminated for data contamination concerns.

VerilogEval: evaluating large language models for Verilog code generation mentions → 90% 1e evidence

The paper cites VerilogEval as a benchmark for RTL code generation.

RTLLM: an open-source benchmark for design RTL generation with large language model mentions → 90% 1e evidence

The paper cites RTLLM as a related benchmark for RTL code generation.

VeriDebug mentions → 90% 1e evidence

The paper cites VeriDebug as related work on bug insertion for dataset building.