CovR: Coverage-Aware Hardware Verification via Reasoning-Guided Reinforcement Learning

Manar Abdelatty; Maryam Nouh; Sherief Reda

cs.ARSep 15, 2026

CovR: Coverage-Aware Hardware Verification via Reasoning-Guided Reinforcement Learning

Authors: Manar Abdelatty, Maryam Nouh, Sherief Reda

Organizations: School of Engineering Brown University Providence, RI, USA

Abstract

Design verification remains one of the most resource-intensive stages of hardware development, often consuming up to 70% of the total design effort. While recent work has explored using Large Language Models (LLMs) to automate testbench generation, most existing approaches focus narrowly on functional correctness, overlooking the critical aspect of coverage quality. To bridge this gap, we present CovR, an agentic framework for automated testbench generation that combines self-reflection loops with simulation-based feedback to maximize coverage. Using this pipeline, we construct a large-scale dataset of 16,514 natural specification RTL reasoning testbench tuples with a strong teacher model, enabling coverage-aware supervision. Building on this, we propose a reinforcement learning (RL) framework tailored for coverage-driven testbench generation, leveraging tool-derived rewards from simulation and coverage feedback to optimize a student model. Experimental results show that the CovR finetuned model achieves 93.81% cov@10 on VerilogEval and RTLLM V2.0, and 87.76% cov@10 on CVDP, outperforming state-of-the-art approaches by 7.97% and 3.59%, respectively. Furthermore, deploying the finetuned model back into the agentic refinement pipeline further improves cov@10 to 94.27% on VerilogEval and RTLLM V2.0 and 91.39% on CVDP. Moreover, when integrated as a plug-in stimulus engine for full verification workflows, CovR improves coverage by 18.95% and mutation detection score by 1.19%, while revealing 4.46% undetected failures, highlighting the importance of optimizing for coverage in LLM-based hardware verification.

Figures & tables

Figure 1. Overview of CovR within a full RTL verification workflow. CovR generates high-coverage stimuli that exercise the DUT, oracle, and injected RTL mutants to improve coverage, mutation detection, and bug-finding effectiveness.

Approach	Coverage Feedback	Agentic	Reasoning	Learning	Primary Output
AutoBench ( Qiu et al., 2024a )	$\times$	$\checkmark$	$\times$	None	FRM + Stimuli Testbench
CorrectBench ( Qiu et al., 2025 )	$\times$	$\checkmark$	$\times$	None	FRM + Stimuli Testbench
PRO-V-R1 ( Zhao et al., 2025 )	$\times$	$\checkmark$	$\checkmark$	SFT-W/Reason + RL	FRM + Test Stimuli
Testplan Gen. ( Kochar et al., 2026 )	$\times$	$\checkmark$	$\checkmark$	SFT-W/Reason + RL	Test-plan
LLM4DV ( Zhang et al., 2025c )	$\checkmark$	$\checkmark$	$\times$	None	Test stimuli
Test stimuli gen. ( Park et al., 2025 )	$\checkmark$	$\times$	$\times$	SFT + RL	Test stimuli (Tabular)

Table 1. Prior work leverages coverage feedback but rarely integrates reasoning with reinforcement learning. CovR unifies these components to produce high-coverage testbenches.

Figure 2. Overview of the CovR framework. (a) Self-refinement workflow: An initial testbench is generated from the specification $\mathcal{N}$ and RTL $\mathcal{V}$ , then iteratively refined through execution and coverage-driven loops to produce a valid high-coverage testbench $\mathcal{T}^{*}$ , which is annotated with reasoning traces $\mathcal{R}^{*}$ . (b) Agentic training framework: The resulting tuples $(\mathcal{N},\mathcal{V},\mathcal{R}^{*},\mathcal{T}^{*})$ form a synthetic dataset for supervised fine-tuning (SFT), followed by reinforcement learning (RL), where multiple testbenches are sampled, simulated in parallel, and optimized using hierarchical rewards combining format, compile, and coverage signals.

Source	Designs	Statistics {Median, Max}
	Count	I/O Ports	# Cells	Branches	FSM Trans.
Pyra ( Nadimi et al., 2025a )	1,732	{7,3586}	{1,16678}	{2,274}	{9,68}
VeriThoughts ( Yubeaton et al., 2025 )	8,775	{12,21276}	{6,3129}	{5,527}	{6,66}
VeriReason ∗ ( Wang et al., 2025b )	998	{8,256}	{3,33046}	{3,80}	{6,6}
VeriTriplets ∗ ( Wang et al., 2025a )	2,309	{54,16772}	{36,36488}	{16,1151}	{7,112}
LLM4Cov ∗ ( Zhang et al., 2026a )	3,468	{30,38452}	{17,47438}	{8,4033}	{7,35}

Table 2. RTL and Testbench design sources.

Figure 3. CovR synthetic dataset coverage distribution. Line, toggle, condition, and branch coverage achieve 100% median, while FSM coverage is more dispersed with a 75% median.

	Model	Variant	Agentic	VerilogEval + RTLLM V2.0						CVDP (cid012)
				pass@k (%)			cov@k (%)			pass@k (%)			cov@k (%)
				k=1	k=5	k=10	k=1	k=5	k=10	k=1	k=5	k=10	k=1	k=5	k=10
Vanilla	gpt-4o-mini	–	$\times$	88.62	95.25	96.06	81.27	90.01	91.60	71.08	87.15	90.30	62.10	79.78	83.75
	Qwen2.5-Coder-7B-Instruct	–	$\times$	77.88	93.72	95.07	67.26	86.54	89.54	51.44	75.22	80.72	41.63	65.05	71.49
	Qwen3-4B-Instruct-2507	–	$\times$	65.02	93.47	96.06	55.56	86.13	90.24	20.00	53.60	69.88	13.82	38.54	51.58
Iterative Feedback	Qwen2.5-Coder-7B-Instruct	–	$\checkmark$	80.30	90.61	92.12	73.70	84.99	87.08	69.64	82.15	84.34	56.72	70.01	73.01

Table 3. CovR versus baseline methods. Evaluation is performed on VerilogEval ( Liu et al., 2023 ) , RTLLM V2.0 ( Lu et al., 2024 ) , and CVDP ( Pinckney et al., 2025 ) . pass@k measures simulation validity, and cov@k reports the average best coverage. Agentic column indicates the inference mode:

\times

denotes direct model inference without tool feedback, while

\checkmark

denotes agentic inference with iterative coverage feedback.

Panel of Figure 4 — Figure 4. Distribution of per-problem pass@ $1$ and cov@ $1$ scores for Qwen3-4B-Instruct-2507 across (a) the vanilla, (b) reasoning-augmented supervised fine-tuning, (c) GRPO-based reinforcement learning optimized models, and (d) the full agentic pipeline.

	Model	Setting	VerilogEval + RTLLM v2.0				CVDP (cid012)
			Eval0	Eval1	Eval2-100	cov@1	Eval0	Eval1	Eval2-100	cov@1
CorrectBench ( Qiu et al., 2025 )	Qwen2.5-Coder-7B	Base	43.4	36.0	6.9	20.8	9.6	8.4	1.2	1.4
CorrectBench ( Qiu et al., 2025 )	Qwen2.5-Coder-7B	+CovR (Ours)	64.5 $(\uparrow 21.2)$	22.2 (-13.8)	9.4 $(\uparrow 2.5)$	43.1 $(\uparrow 22.3)$	48.2 $(\uparrow 38.6)$	14.5 $(\uparrow 6.0)$	0.0 (-1.2)	25.6 $(\uparrow 24.2)$
PRO-V-R1 ( Zhao et al., 2025 )	PRO-V-R1-Qwen3-8B	Base	93.6	52.2	28.6	67.2	89.5	42.2	1.2	38.2
PRO-V-R1 ( Zhao et al., 2025 )	PRO-V-R1-Qwen3-8B	+CovR (Ours)	94.1 $(\uparrow 0.49)$	48.3 (-3.94)	32.0 $(\uparrow 3.44)$	79.3 $(\uparrow 12.10)$	92.8 $(\uparrow 3.30)$	36.1 (-6.10)	1.2 $(\uparrow 0.00)$	55.4 $(\uparrow 17.20)$

Table 4. CovR as a plug-in stimulus engine for full RTL verification workflows. Integrating CovR improves coverage (cov@1) and mutation detection score (Eval2-100) while preserving simulation validity (Eval0). Lower Eval1 under high coverage stimuli indicates that CovR uncovers previously missed verification bugs.

Figure 5. Cov@1 progression for Qwen3-4B-Instruct across inference variants, broken down by coverage metric.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6. CovR dataset sample, showing high coverage testbench and its corresponding reasoning annotations.

Panel of Figure 7 — Figure 7. Efficient evaluation pipeline of CovR . Problems are distributed across GPUs for LLM inference, stored in JSONL format, and evaluated in parallel using Synopsys VCS to compute pass@k and cov@k metrics.

Panel of Figure 8 — Figure 8. CovR improves mutation detection on Prob106_always_nolatches .

Panel of Figure 9 — Figure 9. CovR reveals an oracle bug in the generated FRM for a byte-enable register. The RTL correctly updates only enabled bytes, while the FRM incorrectly replicates byte values across its internal state.

Panel of Figure 10 — Figure 10. CovR generates longer temporal stimuli that reach the peak turn-around case and expose an RTL bug missed by shorter static stimulus sequences.

Panel of Figure 11 — Figure 11. CovR generates temporally diverse stimuli that expose an oracle bug in the generated functional reference model, while static input-combination stimuli fail to reveal the mismatch.

Panel of Figure 12 — Figure 12. CovR reveals an oracle bug in the generated FRM for Prob115_shift18 . The RTL correctly sign-extends during arithmetic right shift by 8, while the FRM constructs an incorrect sign-extension pattern.

Panel of Figure 13 — Figure 13. CovR improves mutation detection on Prob052_gates100 . Pro-V’s random vectors kill parity and obvious stuck-output mutants, but miss reduction-specific corner cases. CovR adds all-zero, all-one, alternating, and one-hot stimuli, improving mutation score from 6/10 to 9/10.

Explore similar work

LLM4Cov: Execution-Aware Agentic Learning for High-coverage Testbench Generation
Feb 18, 2026Hejia Zhang, Zhongming Yu, Chia-Tung Ho +3LLM Agent Training Automated Test Generation
University of California San Diego, USA · NVIDIA · Agentrys
Exploring LLM-based Verilog Code Generation with Data-Efficient Fine-Tuning and Testbench Automation
Apr 16, 2026Mu-Chi Chen, Po-Hsuan Huang, Yu-Hung Kao +6LLM Fine-Tuning Code Generation
Academia Sinica, Taipei, Taiwan · National Taiwan University, Taipei, Taiwan · National Cheng Kung University, Tainan, Taiwan
A Survey on LLM-Integrated Hardware Design Verification
Oct 6, 2026Hao Zheng, Jaime Rafael Imperial, Bardia Nadimi +1LLM Agent Verification Formal Hardware Verification
Bellini College of Cybersecurity, AI, and Computing, University of South Florida Tampa, FL, USA

Feb 18, 2026J/K move · Enter open · S save

LLM4Cov: Execution-Aware Agentic Learning for High-coverage Testbench Generation

Hejia Zhang, Zhongming Yu, Chia-Tung Ho, Haoxing Ren+2

University of California San Diego, USA · NVIDIA · Agentrys

LLM Agent Training Automated Test Generation Agentic RL

Open paper AskPDF arXiv

Execution-aware LLM agents offer a promising paradigm for learning from tool feedback, but such feedback can be expensive and slow to obtain, making online reinforcement learning (RL) less practical in certain scenarios. High-coverage hardware verification exemplifies this challenge due to its reliance on industrial simulators and non-differentiable execution signals. We propose LLM4Cov, an offline agent-learning framework that models verification as single-step state transitions guided by deterministic evaluators. Building on this formulation, we introduce execution-validated data curation, policy-aware agentic data synthesis, and worst-state-prioritized sampling to enable scalable learning under execution constraints. We further curate a reality-aligned benchmark adapted from an existing verification suite through a revised evaluation protocol. Using the proposed pipeline, a compact 4B-parameter model achieves 69.2% pass rate and 90.4% average coverage in CVDP-ECov under agentic evaluation, outperforming its teacher by 5.3% and 10.5%, demonstrating competitive performance against models an order of magnitude larger.

Abstract

Explore similar work

LLM4Cov: Execution-Aware Agentic Learning for High-coverage Testbench Generation

Exploring LLM-based Verilog Code Generation with Data-Efficient Fine-Tuning and Testbench Automation

A Survey on LLM-Integrated Hardware Design Verification