Ziyang Liu

Multimodal Sentiment Analysis Project, School of Computer Science and Technology, Soochow University (Project Leader, 2024.02-2025.03)

• Research Focus: Utilizing multimodal data (text, video, audio) for sentiment analysis, with special attention to robustness improvement in cases of partial modality missing

• Key Contribution: Proposed the Factor-based Semantic Recovery Framework (FSRF), achieving SOTA performance on multiple benchmark datasets

• Applications: Social media sentiment analysis, intelligent customer service, user experience analysis

Multi-Agent Reinforcement Learning Project, Department of Electronic Engineering, Tsinghua University (Core Member, 2025.04-2025.06)

• Project Description: Focus on collaborative optimization and reinforcement learning of multi-agent systems under complex tasks

• Main Work: Construct multi-agent interaction framework, design task environments and implement coordination mechanisms between agents

• Technical Contribution: Combined Test-Time Reinforcement Learning (TTRL) for policy training, effectively improving system generalization performance

• Research Achievements: Independently completed code module development and debugging, assisted in structural analysis and convergence derivation of TTRL theoretical mechanisms

Shenzhen X Institute Pi Project: Self-Referential Iterative Agents (Core Member, 2025.08 - Present)

• Project Content: Exploration of novel agent system paradigms based on self-referential iteration and co-evolution mechanisms, aiming to break through the limitations of existing large language models and agent systems in terms of generality, adaptability, and scalability

• Technical Contribution: Extended the Google DeepMind AlphaEvolve framework by introducing self-referential evolution mechanisms, enabling the system to maintain adaptability while possessing continuous self-improvement capabilities

• Experimental Results: Achieved performance comparable to or better than AlphaEvolve on multiple benchmark tests, while realizing significant computational efficiency improvements at scale

Student Innovation Project: HeartSync - AI-Powered Emotional Support System for Left-behind Children (Project Leader, 2024.06-2024.12)

• Project Overview: Development of AI-based personalized emotional support and psychological counseling system for left-behind children

• Achievements: Won the 2024 X Institute Excellence Project Award and National College Student Innovation Competition Third Prize

AI Memory — AI-Assisted Memoir Writing (Project Lead, 2025.02–2025.04)

Partners: ByteDance Foundation and Accenture.

• Built a memoir-writing prototype for older adults and families, connecting voice input, conversational follow-up, narrative generation, editing, and sharing.

• Implemented ASR/LLM integration, frontend/backend services and storage, with review, correction and export workflows. Large-scale user testing has not been conducted.

Project demo

Automatic Quant Agent — Self-Evolving Quantitative Trading Agent (2026.02–Present)

Partners: Beijing Zhongguancun Academy and Beijing Fund Town.

• Connected data ingestion, factor discovery, portfolio construction and simulated trading with transaction costs, position constraints and risk checks. Simulations covering 2024 and 2025 each achieved cumulative returns above 10%; these are simulated, not live or annualized returns.

• Designed evaluation-driven evolution of factors and the quantitative research framework, with a dashboard for candidate factors, experiments and version tracking.

Orion — Self-Improving Polymarket Trading Agent (Personal Project, 2026.02–2026.04; ongoing iteration)

• Combined market data, news and event context with multi-agent probability forecasting, risk assessment and trade execution, achieving positive returns in personal live-trading validation.

• Used outcome evaluation and supervision to guide component improvements, with a dashboard for trades and experiments. The public repository below focuses on research, risk checks and simulated execution, distinct from the personal live-trading setup.

Project demo / Source code