SiliconMind Project

AI🔄️IC

Project Goals

• Aims to contribute to the expansion of Taiwan’s talent pools in AI model development, data curation, and system implementations

• Initial application will be the use of Large Language Models (LLMs) and Reinforcement Learning (RL) in chip designs

Project Objectives

Develops next-generation reasoning LLMs broadly applicable to scientific research and educational scenarios.

1. Develop highly efficient adversarial reinforcement learning methods to reduce resource requirements while enhancing the reasoning performance and accessibility of LLMs.

2. Enhance LLMs' higher-order abstract reasoning and generalization capabilities, enabling them to tackle complex scientific and educational problems.

3. Develop human-readable and explainable reasoning processes to improve the transparency of model reasoning and facilitate human-AI collaboration.

4. Construct a knowledge reasoning framework suitable for localization and cross-disciplinary applications, thereby lowering the technical barrier for professionals utilizing reasoning LLMs.

5. Leverage the scientific databases and domain expertise of Academia Sinica to build LLM tools that are widely applicable in science education and research.

Anticipated Outcomes

Introduce two application cases leveraging Taiwan's local advantages for practical validation.

1. Key Electronic Design Automation (EDA) tasks in Taiwan's semiconductor industry, to verify the model's capability to handle specialized and highly complex tasks.

2. A Traditional Chinese-centric multilingual and multimodal Large Language Model (LLM), to evaluate the model's abstraction and generalization performance across language and multimodal integration tasks.

Team Members

• PI: Chien-Yao Wang, Academia Sinica

• Co-PIs

  • Academia Sinica: Hong-Yuan Mark Liao, Ai-Chun Pang, Ling-Jyh Chen, Yuan-Hao Chang, Yu Tsao, Li Su, De-Nian Yang, Hen-Hsen Huang, Ti-Rong Wu

  • Universities:

    • National Taiwan University: Shih-Hao Hung, Ming-Syan Chen, Chia-Hsiang Yang

    • National Yang Ming Chiao Tung University: Chen-Yi Lee

    • National Cheng Kung University: Lih-Yih Chiou

    • National University of Kaohsiung: Chun-Hsin Wu

    • National Kaohsiung University of Science and Technology: Yeong-Chau Kuo

• Consultants: James C. Liao, Academia Sinica; HT Kung, Harvard University; Yuh-Jye Lee, Academia Sinica

Project Links

• SiliconMind V1

• Verilog Online Judge

• Kuwa IC Design

計畫目標

1. 協助擴充台灣在 AI 模型開發、資料整備與系統實作等領域的人才庫

2. 初步將把大型語言模型 (LLMs) 與強化學習 (RL) 應用於晶片設計領域

計畫方向

開發能夠廣泛應用於科學研究與教育場景的新一代推理型 LLM

1. 發展高效率的對抗式強化學習方法，降低資源需求並提升LLM的推理性能與可及性

2. 提升LLM處理更高階抽象推理與泛化能力，使之能應對複雜的科學與教育問題

3. 開發人類可讀且具解釋性的推理過程，以提高模型推理透明度並促進人機協作

4. 建構適合本地化與跨學科之知識推理框架，降低專業人員使用推理型 LLM的技術門檻

5. 透過中央研究院之科學資料庫與專業知識，打造可廣泛應用於科學教育與研究之 LLM 工具

實務驗證

導入兩個具有台灣在地優勢的應用案例進行實務驗證

1. 台灣半導體產業重要的電子設計自動化(EDA)任務，以驗證模型處理專業、高難度任務的能力

2. 以繁體中文為主的多語言多模態大型語言模型(LLM)，以評估模型對語言及多模態整合任務的抽象與泛化效能

計畫成員

• 計畫主持人：王建堯 

• 計畫共同主持人

  • 中央研究院資訊科學研究所、資訊科技創新研究中心：
    廖弘源、逄愛君、陳伶志、張原豪、曹昱、蘇黎、楊得年、黃瀚萱、吳廸融

  • 大學：洪士灝(台大)、陳銘憲(台大)、李鎮宜(陽明交大)、邱瀝毅(成大)、楊家驤(台大)、吳俊興(高大)、郭永超(高科大)

• 計畫顧問：廖俊智、孔祥重、李育杰

專案連結

• SiliconMind V1

• Verilog Online Judge

• Kuwa IC Design

Since Feb, 2025