Date of Award

5-5-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Department

Computer Science and Engineering

First Advisor

Xiao Li

Abstract

We built a browser-based Street Fighter III benchmark for comparing large language models (LLMs) and visionlanguage models (VLMs) under real-time constraints. The game runs through a Gymnasium-like control loop that exposes both frames and structured state while the agent emits button actions. We deploy the system on Modal and stream matches overWebRTC, which lets us compare ability, latency, and cost in one setting instead of spreading them across unrelated benchmarks. Because the benchmark is narrow, architectural tradeoffs that are easy to blur in offline evaluation show up quickly during live play.

Share

COinS