Ziyi (Iggy) Zhao

One of my recent works has been accepted for publication in the Journal of the American Statistical Association, the premier journal in statistics and data science. In this paper, we introduce a unified diagnostics framework for generalized linear models—one of the most widely used classes of statistical models in data analysis. Whether you're working with binary data, ordinal data (such as rating scales), or count data, this tool can support your model development and evaluation efforts. Our framework functions like a “CT scan” for your model, offering deep insights into model structure and explainability. https://lnkd.in/dZgYsCDj #GLM #ExplainableAI #ModelTransparency #BusinessAnalytics #DataAnalysis

222

21 Comments

Thrilled to share that I contributed to three papers that were accepted and presented at INFORMS 2025 in Atlanta! 🎉 📘 Text Embedding Models Can Be Great Data Engineers — presented by Dr Murat Yildirim We explored how text embedding models can act as powerful data engineers by converting unstructured information into structured representations ready for optimization and analytics. 📘 Towards a Foundation Model for Sensor-to-Decision Optimization in Energy Systems: An End-to-End Framework with Attention Networks — presented by Mohammad Badfar We proposed an attention-based foundation framework that connects sensor data directly to decision-making in energy systems, enabling end-to-end optimization. 📘 Attention Is All You Need to Optimize Wind Farm Operations and Maintenance — presented by me Our model integrates sensor-driven degradation predictions with multi-head attention to optimize wind farm operations and maintenance planning. It was an amazing experience sharing our research and connecting with the community at INFORMS 2025 in Atlanta! 🚀

63

5 Comments

New ICLR 2026 paper out! We study internal planning in language models by analyzing hidden states through an information-theoretic lens. We study how language models plan ahead by measuring (1) how far into the future their hidden states carry signal, and (2) whether they still represent multiple plausible next steps before committing to one. Thanks to my co-first author Baris Askin and our advisors Gauri J., Carlee Joe-Wong and Guannan Qu. Paper link: https://lnkd.in/eSqjyJEk #ICLR2026, #LLM, #Interpretability

27

Happy to share that our paper, co-authored with Prof. Marc Fredette, has been published in the Findings of EMNLP 2025. We explored a simple question: is testing a Large Language Model with a single answer enough to truly understand its capabilities? Given their high response entropy, we felt a more robust method was needed. Our work introduces a Monte-Carlo framework, inspired by behavioral sciences, to evaluate LLMs with more statistical rigor. We used it to test for cognitive biases and found that newer, larger models are surprisingly more susceptible to them, making their responses more human-like but less classically rational. This emerging area of behavioral analysis for LLMs feels critical for understanding their reliability in various social sciences settings. If you're also working in this space or are interested in the intersection of cognitive science and AI, I'd be very keen to connect and discuss potential collaborations. You can read the paper here:  https://lnkd.in/etvetQqf

59

14 Comments

Check out our new ICLR 2026 paper on understanding how language models plan ahead before emitting the next token, where we analyze the model's hidden states from an information-theoretic perspective. This work was co-led by amazing PhD students Baris Askin and Muhammed Ustaömeroğlu, and in collaboration with my colleagues Guannan Qu and Carlee Joe-Wong. Paper link: https://lnkd.in/ePFa69yW

86

1 Comment

📊 Tutorial Spotlight at Interspeech 2025 Extracting Insights from Your Complex Data: Interpretable Statistical Methods in Speech Science Presented by Tyson Barrett, Tristan J. Mahr, Visar Berisha, and Camille Wynne Speech science data is often complex—but your analysis doesn’t have to be. This tutorial introduces interpretable statistical methods tailored for speech researchers and clinicians who want clear insights from messy or high-dimensional data. 🔍 Topics include: Causal inference & statistical modeling Bayesian inference made intuitive Linear regression & mixed effects models Generalized additive models (GAMs) Real-world examples and case studies No advanced math required—just a basic foundation in stats and some R or Python experience. This is the tutorial every speech researcher wished they had early on. Perfect for those working at the intersection of data and speech science who want to level up their analytical skills. 📅 Join us in Rotterdam at #Interspeech2025! 🔗 https://lnkd.in/eFTnzXNb

13