From Mirage to Grounding: Towards Reliable Multimodal Circuit-to-Verilog Code Generation

2026-04-30

AI Summary

This work addresses the “mirage” problem in multimodal large language models for circuit diagram-to-Verilog code generation, where models over-rely on module identifiers while neglecting visual content, leading to unreliable outputs. The study is the first to identify and quantify this issue, introducing C2VEVAL—a new evaluation benchmark that employs Normal and Anony protocols to assess models’ genuine dependence on visual information. The authors propose VeriGround, a 4B-parameter multimodal model enhanced with identifier anonymization, rejection-sample augmentation, and decision-focused D-ORPO preference optimization to precisely control critical tokens. Experiments show that VeriGround achieves functional Pass@1 rates of 46.11% and 42.51% under Normal and Anony settings, respectively, with false rejection rates as low as 1.20% and 0.00%, and rejects over 92% of blank-image inputs—matching GPT-5.4’s performance and significantly outperforming existing baselines.

Abstract

Multimodal large language models (MLLMs) are increasingly used to translate visual artifacts into code, from UI mockups into HTML to scientific plots into Python scripts. A circuit diagram can be viewed as a visual domain-specific language for hardware: it encodes timing, topology, and bit level semantics that are invisible to casual inspection yet safety critical once fabricated in silicon. Translating such diagrams into register-transfer-level(RTL) code therefore represents an extreme reliability test for vision-to-code generation. We reveal a phenomenon we call Mirage: replacing a circuit diagram with a blank image leaves Pass@k unchanged or even higher, because models bypass the visual input and instead exploit identifier semantics in the module header to retrieve canonical RTL templates. This constitutes a new, highly covert class of defect in AI-assisted code generation that directly undermines MLLMs' trustworthiness. To quantify the effect, we construct C2VEVAL and evaluate eight MLLMs under a paired Normal/Anony protocol in which Anony mode anonymizes all identifiers in both the diagram and the module header; Anony-mode scores drop sharply across all models, confirming that high Normal-mode accuracy is largely a Mirage. We then propose VeriGround (4B), trained with identifier anonymization, refusal augmentation, and D-ORPO (Decision-Focused ORPO) preference alignment that up-weights pivotal generate-or-refuse tokens. VeriGround achieves Functional Pass@1 of 46.11%/42.51%(Normal/Anony) with a False Refusal Rate of only 1.20%/0.00%, while maintaining >92% Refusal Rate on blank images. With only 4B parameters, VeriGround performs on par with GPT-5.4 under Normal and significantly outperforms all baselines under Anony, confirming genuine visual grounding.

Problem

• Multimodal LLMs
• Circuit-to-Verilog
• Visual Grounding
• Mirage Effect
• RTL Code Generation

Innovation

• Mirage phenomenon
• visual grounding
• identifier anonymization
• D-ORPO
• multimodal code generation