School of Engineering Department of Computer Science and Engineering 125 An LLM-Based Multi-Agent System for Financial Trading Simulation and Strategy Backtesting Supervisor: GUO Song / CSE Co-Supervisor: ZHANG Jie / CSE Student: CHUNG Ho Man / COGBM LI Yang / ECOF LAI Sing Kwong / COMP Course: UROP 1000, Summer UROP 1100, Summer UROP 1000, Summer Existing LLM-based financial systems struggle with challenges in extreme market scenarios, cross-market knowledge transfer, agent coordination, and self-learning from failures. To address these limitations, we introduce FinThink, a multi-agent system with a cognitively chained architecture. Its core contributions include: 1) the STL Protocol, which converts unstructured sentiment into structured signals; 2) a Hierarchical Reusable Memory System that achieves self-correction by analyzing past decisions and generating corrective strategies; 3) a Unified Coordination Framework that orchestrates agent collaboration via a finite-state routing engine; and 4) a Market Cognition Settlement System that provides comprehensive analysis by combining short-term and long-term perspectives. We will continue to optimize memory algorithms, balance multi-temporal cognition, and enhance the system’s ability to adapt to dynamic markets. Decoupled KV Cache Compression for Efficient Long-Context LLM Inference Supervisor: GUO Song / CSE Student: HO Lok Yin / COMP Course: UROP 1100, Summer This progress report summarizes initial exploration toward decoupled KV cache compression for efficient long-context LLM inference. Motivated by the high memory and bandwidth demands of KV caches in transformer decoding, the project investigates compressing keys and values separately according to their differing statistical properties. Using nanoGPT on the Tiny Shakespeare dataset, I conducted exploratory trials with simulated low-bit quantization (FP4, FP8, INT4, INT8) and bitsandbytes-based quantization (NF4, INT4, INT8) to study trade-offs in memory footprint, decoding speed, and perplexity. Prior to these experiments, a literature review was conducted covering recent KV cache optimization methods, quantization strategies, alternative architectures, and system-level approaches, providing the theoretical grounding for the experimental design. Foundation Model-as-a-Service at Edge Supervisor: GUO Song / CSE Student: YANG Yuhan / MATH-STAT Course: UROP 1000, Summer This report outlines the progress made during the summer on research into machine learning applications for multi-agent systems. I completed two foundational courses in machine learning, covering supervised and unsupervised learning techniques, and initiated studies on reinforcement learning. Furthermore, I conducted a review of cutting-edge research papers on multi-agent reinforcement learning, with a focus on coordination, communication, and learning challenges within decentralized systems. This work has enhanced my understanding of key concepts such as credit assignment and exploration in multi-agent systems. Future plans include applying the acquired theoretical knowledge to practical research and development, specifically by constructing multi-agent systems simulations using Python to implement and validate the reviewed algorithms.
RkJQdWJsaXNoZXIy NDk5Njg=