Chao Péter Yang
Machine Learning Researcher
Yale University, Department of Computer Science
Duke Interpretable Machine Learning Lab
Biography
Chao Péter Yang is a machine learning researcher working on LLM alignment, agentic systems, and structure-aware generative modeling. He is a Research M.S. student in Computer Science at Yale University, advised by Prof. Rex Ying, and holds an M.S. in Data Science from Duke University, where he was advised by Prof. Cynthia Rudin, Dr. Stephen Ni-Hahn, and Prof. Brandon Fain. He is a first author at NeurIPS 2025 for ProGress, a structured symbolic music generator combining discrete diffusion with hierarchical music analysis, and his recent work audits and mitigates sycophancy and agreement bias in preference-aligned language models.
Previously, he was a Data Scientist Intern at Amazon Robotics, where he built an AI agent for warehouse root-cause investigation, and spent three years deploying production ML in financial analytics as a Senior Data Scientist at Curinos, Inc. He graduated with Highest Honors in Data Science from the University of Michigan and has served as a reviewer for NeurIPS and KDD.
Interests
- LLM Alignment (DPO / RLHF)
- Sycophancy & Agreement Bias
- Agentic & Multi-Agent Systems
- Structure-Aware Generative Modeling
- Symbolic Music Generation
- Graph Neural Networks
- Interpretable Machine Learning
Education
Research M.S. in Computer Science (Thesis Track), 2026 - 2028
Yale University
M.S. in Data Science, 2024 - 2026
Duke University
B.S. (Hon.) in Data Science and Mathematics, Minor in Music, 2018 - 2021
University of Michigan - Ann Arbor
International Baccalaureate, 2018
American International School of Budapest
Experience
Research Assistant
- Audited sycophancy at every post-training stage of open-recipe LLMs (OLMo 3, Tülu 3, SmolLM3), showing it is largely introduced by SFT rather than DPO; first-author draft targeting FAccT 2027.
- Built sycotrace, an interpretability pipeline that extracts and steers a sycophancy direction in activation space (persona vectors, CAA) and scores training data along it; found the direction already present in a pretraining-only 3B model.
- Developed PrefAug, an audited preference-data augmentation pipeline, with controlled TRL DPO/LoRA ablations on OLMo-3 and Qwen3.5-9B; engineered the full PyTorch/SLURM stack with LLM-judge evaluation and W&B logging.
- Advised by Prof. Brandon Fain.
Research Assistant
- Built ProGress, a structured symbolic music generator combining discrete diffusion with a knowledge-based probabilistic model; a 45-subject study beat baselines on quality with only ~3M parameters. Accepted to NeurIPS 2025.
- Implemented a custom DiffPool for a heterogeneous GNN used in musical analysis (PyTorch), reducing validation cross-entropy loss by 60% via architecture and hyperparameter tuning.
- Developed a Group Relative Policy Optimization (GRPO) algorithm over a Graph Neural Network to bring RLHF to automated, personalized musical analysis.
- Advised by Prof. Cynthia Rudin and Dr. Stephen Ni-Hahn.
Data Scientist Intern
- Researched and built an AI agent for warehouse root-cause investigation that integrates multiple data sources and MCP servers, cutting average troubleshooting time from several days to 2.5 minutes at a 75% success rate.
- Built a reusable agentic framework on LangGraph and Amazon Bedrock that standardized internal agent development and shortened the path from prototype to deployment.
- Engineered a production evaluation pipeline for large-scale agent benchmarking using LLM-as-a-Judge with Langfuse, enabling rapid, repeatable performance measurement.
Senior Data Scientist, Modeling Research
Curinos, Inc. (Informa PLC subsidiary) – Banking Consulting
- Researched and developed novel, industry-grade nonlinear elasticity models for Asset–Liability Management (ALM), improving out-of-sample R² over legacy models by ≈12% across bank and credit-union portfolios.
- Created automated ad-hoc regression notebooks with PySpark for creating, testing, and validating models with different configurations, reducing the time to build proof-of-concept models by half.
Data Scientist II, Modeling Research
Curinos, Inc. (Informa PLC subsidiary) – Banking Consulting
- Led an ML-engineering team migrating the legacy modeling pipeline from Cloudera to Databricks, coordinating testing, promotion, and release across teams, saving $100k+ annually and cutting average data-processing time by 30%; recognized at a company-wide town hall.
- Tuned hierarchical nonlinear price-elasticity models en masse for major US banks (10,000+ segments each), improving AIC and R² with markedly higher convergence rates.
- Installed and managed more than 10,000 price elasticity models per client bank to predict and optimize their deposit portfolio across a wide range of interest rates, with precise Model Risk Management documentation.
Data Scientist, Modeling Research
Curinos, Inc. (Informa PLC subsidiary) – Banking Consulting
- Converted the local, single-threaded legacy modeling pipeline to use SparkR and Cloudera, reducing run time for model fitting by up to 30x.
- Performed Exploratory Data Analysis (EDA) for client banks to tune and reconfigure their models and data segments, leading to better-performing price elasticity models in terms of MAPE, R², and rate of convergence.
- Set up and automated custom SQL procedures to clean, wrangle, map, and transform client data feeds for the modeling pipeline, partially eliminating the need for manual model data refreshes.
Honors Student Researcher
- Researched content-based music classification with neural networks, advised by Prof. Edward Ionides and Prof. Daniel Forger.
- Developed new music classification methods using MIDI and LSTM networks, reaching 82% accuracy, more than 10% above conventional ML methods.
- Received the Highest Honors distinction in Data Science, one of only two awarded in 2021.
Awards & Honors
Full Scholarship + Stipend
Research M.S. in Computer Science, Yale University
Outstanding Capstone Award
ClearPath Agent capstone selected as 1 of 2 outstanding projects for bridging industry and academia
Dean’s Research Award
MIDS Merit Scholarship
70% merit scholarship, M.S. in Data Science
Highest Honors in Data Science
One of two awarded in the department; University Honors (2019, 2021)
Service
Reviewer: NeurIPS 2025 AI4Music Workshop · NeurIPS 2026 AI4Music Workshop · NeurIPS 2026 Creative AI Track · KDD 2026 AI4Sciences Track · KDD 2027 AI4Sciences Track
Certificates
Gain foundational knowledge, practical skills, and a functional understanding of how generative AI works
See certificate
Introduction to Scala
Deep Learning Spcialization
Share Data Through the Art of Visualization
Projects
*
A multi-agent system that replicates human investment committees with a Supervisor coordinating Research, Quant, and Risk agents via ReAct, improving reasoning accuracy and risk awareness by up to +16% over consensus baselines. Awarded Outstanding Capstone (1 of 2).
A healthcare data dashboard integrated with a fine-tuned LLM-powered chatbot, enabling dynamic querying, interactive visualizations, and scalable cloud deployment.
An AI-powered platform that helps Duke students find professors whose research aligns with their interests, using natural language queries and graph-based exploration.
A research project into developing a model that can create scores from pieces of music.
Performed EDA and a simple XGBoost to predict house prices in California in a single Jupiter notebook. This is simple data project to showcase how I’d approach a relatively straight forward modeling task.
Developing use cases and documentation for the Squirrels API
Publications
Quickly discover relevant content by filtering publications.
(2026).
Mitigating Agreement Bias in LLMs via Contradictory-Knowledge Prompt Triads and Reinforcement Training.
In preparation.
(2026).
FinCom: A Financial Multi-Agent Demo with Disagree-or-Commit Deliberation.
arXiv 2026.
(2026).
Getting Motif-ated: Controllable AI Compositions from Injected Motif Prompts.
NeurIPS 2026 (submitted).
(2026).
SchenkerLink: Human-in-the-Loop Hierarchical Music Analysis with Uncertainty-Aware Graph Link Prediction.
KDD 2027 (under review).
(2025).
ProGress: Structured Music Generation via Graph Diffusion and Hierarchical Music Analysis.
NeurIPS 2025.
(2021).
The Classical-Romantic Dichotomy: A Machine Learning Approach.
Honors Thesis.
Contact
Feel free to leave me a message and I’ll get back to you as soon as possible!
- chaopeter.yang@gmail.com
- New Haven, CT