Event
Ph.D. Research Proposal: Yiting Wang
Tuesday, November 25, 2025
10:00 a.m.
AVW 2460
Sarah Pham
301 473 2449
spham124@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Yiting Wang
Committee:
Professor NAME (Chair) : Gang Qu
Professor NAME: Ang Li
Professor NAME: Cunxi Yu
Date/time: November 25, 2025 from 10 am -1 pm
Location: AVW 2460
Title: Toward Intelligent Automation of LLM-Based Hardware Design
Abstract: This proposal presents a comprehensive Large Language Model (LLM)-based hardware design flow by integrating generation, optimization, and verification. Traditional hardware design requires engineers to manually write RTL code, tune optimization parameters from synthesis feedback, craft Tcl scripts, write verification scripts, and coordinate multiple Electronic Design Automation (EDA) tools —an effort that is highly time-consuming and error-prone. Our research shows how LLMs can serve as the central intelligence orchestrating the design process, producing satisfying results with minimal human intervention. The proposed flow integrates four LLM-powered components: VeriReason – a reinforcement learning–enhanced RTL generation system with reasoning capabilities, achieving 83.1% functional correctness through testbench feedback. SymRTLO – a LLM-driven optimization framework achieving up to 43.9% power, 62.5% timing, and 51.1% area improvements via symbolic reasoning. HADA – an automated hardware security verification system using multi-source data augmentation. MCP4EDA – a Model Context Protocol (MCP)–based backend-aware synthesis TCL script optimizer achieving 15–30% timing and 10–20% area improvements through closed-loop feedback. Experimental results show that this integrated approach not only surpasses fixed EDA scripts in design quality but also drastically reduces development time and manual effort. This work establishes a new paradigm in hardware design automation, where intelligent LLM agents collaborate with EDA tools to deliver a seamless, efficient, and democratized path to high-quality hardware.
