Researchers have identified a significant reliability issue in multimodal large language models (MLLMs) when generating hardware description language (HDL) code from circuit diagrams. This "Mirage" phenomenon occurs when models bypass visual input, relying instead on textual identifiers to retrieve pre-existing code templates, leading to high accuracy even with blank images. To address this, a new model called VeriGround (4B) was developed, incorporating techniques like identifier anonymization and a novel preference alignment method (D-ORPO) to ensure genuine visual grounding. VeriGround demonstrates competitive performance, outperforming baselines on anonymized inputs and maintaining a high refusal rate for invalid inputs. AI
Summary written by gemini-2.5-flash-lite from 2 sources. How we write summaries →
IMPACT Highlights a new class of failure in MLLMs for code generation, necessitating more robust evaluation and training methods.
RANK_REASON Academic paper introducing a new phenomenon and a model to address it.