Hi, I'm Haoyu Gu!
Undergraduate Student in Artificial Intelligence
I'm passionate about Music Representation and Generation, Affective Computing, and Brain-Computer Interface. Let's build something interesting together! 🚀
📞 +86 18100607076 (WeChat: same number) | 📧 ghy20050104@gmail.com
📍 Nanjing, Jiangsu, China | Last Update: April 19, 2026
About Me
Current Status
I'm a junior undergraduate student majoring in Artificial Intelligence at the School of Future Technology, South China University of Technology (SCUT).
Location
Based in Nanjing, Jiangsu Province. School located in Guangzhou, Guangdong Province.
Education
South China University of Technology
985 Project University, Excellence 9 (E9) League Member
未来技术学院 - One of China's first 12 Future Technology Schools
Major: Artificial Intelligence
Academic Performance: GPA 3.95+
Core Courses: C++ Programming, Python Programming, Discrete Mathematics, Data Structures, Computer Networks, Computer Organization and Architecture, Database Systems, Circuit Analysis and Analog Circuits, Digital Circuits, Signals and Systems, Digital Signal Processing, Machine Learning, Deep Learning and Computer Vision
Nanjing Ninghai High School
High School Affiliated to Nanjing Normal University, Shuren School
Nanjing Langya Road Primary School
Honors & Awards
Future Technology Taihu Innovation Award
学业创新一等奖
无锡市政府
Lanqiao Cup C++ Programming Contest (Group A)
广东省二等奖
工业和信息化部人才交流中心
China Mathematics Competition
广东赛区二等奖
中国数学会
SCUT Mathematical Contest
一等奖
华南理工大学数学学院
China Undergraduate Mathematical Contest in Modeling
广东省二等奖
中国工业与应用数学学会
Embedded Chip and System Design Competition
南部赛区二等奖
中国电子学会
China Mathematics Competition
广东赛区三等奖
中国数学会
Research Interests
Focused on Arts & Humanities Computing — bridging AI with music, emotion, and the brain.
Music Representation and Generation
Exploring how to represent and generate music with machines. Interested in finding better ways to encode musical structure and enable AI to compose coherent, expressive pieces.
Affective Computing
Interested in how AI can recognize and respond to human emotions. Building systems that understand affective signals across different modalities.
Brain-Computer Interface
Exploring EEG signal processing and neural decoding. Interested in understanding brain activity patterns and building intelligent systems that bridge neural signals with real-world applications.
Research Outputs
Dual-Track Piano Music Generation
AI-generated dual-track piano pieces with coherent structure and expressive harmony
音乐续写：Let It Go
给定《冰雪奇缘》Let It Go 前12秒，生成后续内容
原曲（更有节奏与情感）：
生成（更加连贯和符合乐理）：
长期语义连贯
全曲主题发展连贯，乐句结构清晰，旋律走向符合音乐逻辑
双声部协调
钢琴左右手配合默契，旋律对话清晰，声部走向和谐
创造性表达
和声进行富有新意，旋律转折出人意料，节奏型设计巧妙
调性转换
调式转换自然流畅，使用合理的转调和弦，不同调性之间过渡平滑
Multitrack Ensemble
Multiple instruments in coordination
String Quartet
Two violins, viola, and cello
Piano Quartet
Piano, violin, viola, and cello
Piano & Choir
Piano with SATB choir
Clarinet & Piano
Clarinet with piano accompaniment
Publications
Pianoroll-Event: A Novel Score Representation for Symbolic Music
Proposed a novel Pianoroll-Event representation that converts Pianoroll into event sequences for efficient compression while ensuring lossless information. This method significantly reduces sequence length and improves generation quality. Responsible for core encoding design, main paper writing, and critical code implementation.
Under ReviewBEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps
Symbolic music generation has advanced rapidly with Transformers, yet a fundamental question persists: what constitutes a good representation? The field has split between event-based and piano-roll representations. We identify the information asymmetry in piano-roll as the key obstacle: information is dense along the temporal axis but sparse along the pitch axis. We propose BEAT (Beat-based Encoding with Anchored Tones), which patches along time and tokenizes along pitch, achieving both structural fidelity and sequence efficiency. BEAT enables autoregressive modeling of piano-roll while preserving its structural advantages, demonstrating state-of-the-art performance and unique capabilities including real-time accompaniment generation.
Under ReviewBeatEdit: Symbolic Music Generation as Explicit Editing
Proposed BeatEdit, the first framework for symbolic music generation based on explicit edit operations. Traced the absence of edit-based methods in music to the representational level and identified a beat-level structured encoding as uniquely edit-compatible. Three complementary mechanisms span an axis of increasing edit density: per-token sequence tagging for error correction, iterative refinement for accompaniment editing, and tag-then-fill for segment completion. All share a single encoding and pre-trained backbone, outperforming autoregressive and diffusion baselines while remaining efficient with non-autoregressive parallel inference. Cross-encoding analysis reveals notable encoding-method interaction effects on editing effectiveness.
ACL 2026Anchored Cyclic Generation: A Novel Paradigm for Long-Sequence Symbolic Music Generation
Proposed the Anchored Cyclic Generation (ACG) paradigm that uses anchor features from identified music to guide subsequent generation, effectively mitigating error accumulation in autoregressive models. Introduced the Hierarchical Anchored Cyclic Generation (Hi-ACG) framework with a global-to-local generation strategy and Piano Token representation, achieving superior performance in long-sequence symbolic music generation and demonstrating excellent generalization to music completion tasks.
ACL 2026EmoMM: Benchmarking and Steering MLLM for Multimodal Emotion Recognition under Conflict and Missingness
Introduced EmoMM, a comprehensive benchmark for evaluating Multimodal Large Language Models (MLLM) in emotion recognition under modality conflict and missingness. Uncovered the Video Contribution Collapse (VCC) phenomenon and proposed CHASE (Conflict-aware Head-level Attention Steering), a lightweight mechanism that enhances MLLM reliability in complex affective scenarios without retraining.
Under ReviewNeuroCadence: Controllable EEG-to-Music Generation via Continuous Affective Intermediates
Proposed NeuroCadence, a modular EEG-to-music generation framework that replaces end-to-end black-box mapping with an explicit valence-arousal intermediate, making brain-conditioned music generation interpretable and diagnosable. Introduced a principled diagnostic protocol that localizes error contributions across stages, providing the first systematic attribution of failure modes in affective brain-to-music systems. Explored continuous numeric conditioning as an alternative to text prompts for fine-grained affective control, offering a generalizable paradigm for any brain-conditioned generation system with independently labelable intermediates.
Fractal-BEAT: Beat-Wise Multi-Scale Tokenization for Symbolic Music
Proposed FractalBeat, a resolution-invariant multi-scale beat-level tokenization for symbolic music that replaces hand-crafted pattern encodings with a learned codebook, making the vocabulary size independent of temporal resolution. Enables unified tokenization of quantized score MIDI and expressive performance MIDI within a single vocabulary and language model.
Talks & Presentations
AI Journey: From Study to Research
As a member of the AI Association Academic Resources Department, I was invited to share insights on academic studies, competitions, research, and essential computational tools for computer science majors.
Academic Planning & Experience Sharing
Invited as a student representative to share academic insights and planning strategies with the incoming class of 2025 freshmen.
University Life Guide for New Students
Invited as a student representative to welcome and guide the class of 2024 freshmen, sharing insights on university academic life and providing practical guidance for their college journey.