Zhirong Chen, Kaiyan Chang, Zhuolin Li, Cangyuan Li, Xinyang He, Chujie Chen, Mengdi Wang, Haobo Xu, Yinhe Han, Huawei Li, Ying Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The emergence of multimodal large language models (MLLMs) presents promising opportunities for automation and enhancement in Electronic Design Automation (EDA). However, comprehensively evaluating these models in circuit design remains challenging due to the narrow scope of existing benchmarks. To bridge this gap, we introduce MMCircuitEval, the first multimodal benchmark specifically designed to assess MLLM performance comprehensively across diverse EDA tasks. MMCircuitEval comprises 3614 meticulously curated question-answer (QA) pairs spanning digital and analog circuits across critical EDA stages—ranging from general knowledge and specifications to front-end and back-end design. Derived from textbooks, technical question banks, datasheets, and real-world documentation, each QA pair undergoes rigorous expert review for accuracy and relevance. Our benchmark uniquely categorizes questions by design stage, circuit type, tested abilities (knowledge, comprehension, reasoning, computation), and difficulty level, enabling detailed analysis of model capabilities and limitations. Extensive evaluations reveal significant performance gaps among existing LLMs, particularly in back-end design and complex computations, highlighting the critical need for targeted training datasets and modeling approaches. MMCircuitEval provides a foundational resource for advancing MLLMs in EDA, facilitating their integration into real-world circuit design workflows. Our benchmark is available at https://github.com/cure-lab/MMCircuitEval.

Hardware description language (HDL) code designing is a critical component of the chip design process, requiring substantial engineering and time resources. Recent advancements in large language models (LLMs), such as GPT series, have shown promise in automating HDL code generation. However, current LLM-based approaches face significant challenges in meeting real-world hardware design requirements, particularly in handling complex designs and ensuring code correctness. Our evaluations reveal that the functional correctness rate of LLM-generated HDL code significantly decreases as design complexity increases. In this article, we propose the AutoSilicon framework, which aims to scale up the hardware design capability of LLMs. AutoSilicon incorporates an agent system, which (1) allows for the decomposition of large-scale, complex code design tasks into smaller, simpler tasks; (2) provides a compilation and simulation environment that enables LLMs to compile and test each piece of code it generates; and (3) introduces a series of optimization strategies. Experimental results demonstrate that AutoSilicon can scale hardware designs to projects with code equivalent to over 10,000 tokens. In terms of design quality, it further improves the syntax correctness rate and functional correctness rate compared with approaches that do not employ any extensions. For example, compared to directly generating HDL code using GPT-4-turbo, AutoSilicon enhances the syntax correctness rate by an average of 35.8% and improves functional correctness by an average of 35.6%.

Soft skills, such as personality characteristics, communication skills and leadership, affect personal career performance greatly. Therefore, predicting the soft skills of interviewees can provide interviewers with a strong reference for the decision of hiring. Nowadays, as asynchronous video interviews have gradually become a popular form of interviews, automatic interview evaluation of soft skills has attracted widespread attention from researchers. However, existing automatic evaluation methods have two significant drawbacks. Firstly, most of them model the problem as multi-modal fusion of long-term sequences, while ignoring the consistency of multi-modal expression in short-time windows, which is a key attribute of the interview scene. Secondly, without embedding of professional knowledge in the interview field, the interpretability of the model is relatively weak. To address the above problems, we propose a novelMulti-modal Window-Consistency Fusionnetwork, namely MWCF, to capture the expression consistency of different modalities in a short-time window and re-weight the language signals to enhance important portions in verbal clues. Meanwhile, in order to enhance the interpretability of the evaluation model, we introduce the professional knowledge of interviewers by proposing a topic generation module based on question attention, and embedding the most representative keywords under different soft skills into the model. Furthermore, a real-world interview dataset is built by developing an asynchronous interview platform, and extensive experiments are conducted to show the superior performance of our proposed model.

Traditional person re-identification aims to retrieve the surveillance images containing the same pedestrian. As the quick development of modern cities, a large number of multimodal personal information, including mobile information and social network log, is available and useful for customer identification and crime tracking. Compared with traditional person reidentification (ReID) only based on image modality, how to make full use of multi-modal information for efficient person ReID is more challenging. In this paper, we propose a brand new cross-modal pedestrian retrieval task based on a novel multi-modal dataset containing GPS trajectories and surveillance images. Three sub-tasks are evolved in the benchmark: unsupervised GPS-to-Image, Image-to-GPS, and Image-to-Image retrieval. In order to further verify our ideas, we propose the Similarity Driven Model (SDM), which utilizes the attention mechanism to improve the performance of domain adaptation. Furthermore, we build a cross-modal heterogeneous graph based on SDM, and adopt Triplet-Walk to uniformly represent different modalities for retrieval. Experimental results demonstrate that our method achieves the state-of-the-art on the GPS-ReID dataset.