About
Undergraduate Researcher · University of Washington
I'm Ping Wang (王平), an undergraduate studying computer science at the Paul G. Allen School, University of Washington. I work with Prof. Noah A. Smith at UW and, since May 2026, with Prof. Yi-Hsuan Yang in the Music and AI Lab at National Taiwan University.
My research focuses on making generative models more controllable, adaptive, and aligned with human intentions. I am particularly interested in post-training, reinforcement learning, and multimodal learning: how models can learn from feedback, how to design meaningful reward signals, and how models can adapt through interaction. I am also interested in how these ideas extend to multi-agent systems, where agents must coordinate and respond to one another in dynamic environments.
Much of my recent work explores these questions through music and audio, including rubric-based optimization for text-to-music generation, human-aligned evaluation of music transcription, and audio-to-symbolic music understanding. Through these projects, I aim to develop methods that improve how generative models learn, understand, and interact with complex information, with applications both within and beyond music.
Research interests
- Human-Aligned Generative Models
- Preference-Based Post-Training
- Multimodal Learning and Evaluation
- Reinforcement Learning
- Music and Audio Generation
- Music Information Retrieval
Education
University of Washington
Seattle, WAB.S. in Computer Science, Paul G. Allen School of Computer Science & Engineering · GPA 3.91/4.0
Expected Jun 2027- Relevant coursework: Machine Learning, Natural Language Processing, Data Structures and Algorithms, Distributed Systems, Operating Systems, and Computer Networks.
National Experimental High School at Hsinchu Science Park
Hsinchu, TaiwanHigh School
2020 – 2023
Research experience
Music and AI Lab, National Taiwan University
Taipei, TaiwanStudent Researcher — Advisor: Prof. Yi-Hsuan Yang
May 2026 – Present- SheetSage-Pro: A bilingual, lyric-aligned audio-to-lead-sheet transcription pipeline that extracts melody, chords, beats, meter, and musical structure from full-song audio, with beat tracking, pitch correction, and rhythm quantization evaluated to improve quality. Accepted for presentation at ISMIR 2026 Late-Breaking/Demo.
- Lyrics-to-Song Generation with MIDI-LLM and MIDI-SAG: An end-to-end, lyrics-conditioned song generation pipeline that connects symbolic lead-sheet generation with arrangement and audio synthesis. Fine-tuning language models to generate melody, chords, rhythm, and lyric-aligned symbolic representations, including syllable-level lyric pointers, phonetic information, rests, breaths, and many-to-many lyric–note alignment for controllable generation.
Noah’s ARK, University of Washington
Seattle, WAStudent Researcher — Advisor: Prof. Noah A. Smith
Aug 2025 – Present- A Dual Evaluation for Music Transcription: A dual evaluation framework for notation accuracy and performance fidelity in automatic music transcription: 24 pipelines (eight audio-to-MIDI models × three MIDI-to-score systems) on 230 piano performances, an audio-only listening study with 106 participants and 3,180 pairwise judgments, and a comparison of human preferences with notation-based, signal-processing, and learned audio similarity metrics.
- Rubric-Based Optimization for Text-to-Music Generation: Structured audio-language-model rubric rewards for text-to-music post-training: preference pairs for DPO on MusicGen-small and ACE-Step v1, and scalar rewards with DiffusionNFT on ACE-Step. Rubric rewards gave broad gains on independent perceptual and objective metrics, while specialized rewards worked better for precise musical attributes.
Student Researcher
Sep 2024 – Oct 2025- Embedded Gesture Recognition: Lightweight neural networks for real-time gesture recognition on resource-constrained embedded devices, including model quantization and deployment for on-device inference.
- Tuberculosis Screening from Cough Audio: Transformer-based models, including the Audio Spectrogram Transformer (AST) and MobileViT, for cough-based tuberculosis risk classification.
- Cough Detection and Localization: Machine learning models for detecting cough events and localizing them in time within continuous audio recordings.
Industry experience
Amazon
Taipei, TaiwanSoftware Development Engineer Intern
Jun 2025 – Sep 2025- Embedded firmware development: Designed and implemented firmware for inter-chip communication, camera-module integration, and Wi-Fi connectivity validation on smart-home devices.
- Hardware–software integration: Board-level integration and debugging across embedded hardware, firmware, camera modules, and network interfaces.
- Testing infrastructure: Automated unit tests and peripheral mock frameworks for validating firmware across hardware states, improving test coverage and reducing regression risk.
Phison Electronics
Miaoli, TaiwanMachine Learning / Embedded Systems Engineer Intern
Jun 2024 – Aug 2024- LLM memory modeling: A parameter-calculation system that estimates large language model memory use, including weights and activations, directly from model configuration files.
- Transformer architecture analysis: Analyzed LLaMA 2/3, CodeLlama, Falcon, Mistral, Mixtral, Breeze, Whisper, and Vicuna, focusing on attention architectures, model scaling, and memory requirements.
- Large-scale model training: Worked with multi-GPU Linux infrastructure and Phison AI SSD systems for model training and experimentation.
Skills
- Languages
- Python, C, C++, Java
- Machine learning
- PyTorch, Transformers, preference optimization (DPO), reinforcement learning, diffusion / flow-matching models, LoRA
- Music & audio
- Automatic music transcription, music generation, audio representation learning, symbolic music modeling, MIDI, MusicXML
- Systems & tools
- Linux, Git, distributed training, multi-GPU computing, embedded systems, FPGA