About me

Greetings! I'm Tzu-Heng (Brian), a final-year CS Ph.D. student at the University of Wisconsin-Madison (UW-Madison), working with Frederic Sala in the Sprocket Lab.

I have been fortunate to intern across both industry and national labs. Most recently, I interned at Apple (second time), where I worked on automated verifiers for open-ended tasks in visual reinforcement learning. Previously, I interned at Meta GenAI (now MSL), focusing on synthetic data generation, and earlier at Apple on data curation and data mixing strategies for multimodal pretraining, under the guidance of Javier Movellan and Manjot Bilkhu. In 2019, I was a research intern at Argonne National Laboratory with the Array of Things team, working with Charlie Catlett and Rajesh Sankaran on large-scale urban sensing systems.

Before joining UW–Madison, I earned my B.S. degree in CS from National Chengchi University (NCCU), where I was advised by Man-Kwan Shan and Ling-Jyh Chen. My early research focused on spatio-temporal machine learning and large-scale sensor networks.

My current research focuses on data-centric AI for multimodal models, with the goal of enabling systems to learn more from less but higher-quality supervision. I have worked on several data lifecycle projects, including multimodal data selection, universal data mixing, zero-cost labeling systems, and LLM verification through synthetic programs and rubrics. These efforts are grounded in weak supervision frameworks to build foundation models with fewer human annotations. Additionally, I am exploring a new notion, parameter marketplace, to accelerate training while monetizing parameters as a second profit center.

Recent News

  • We release PAJAMA, a program-as-a-judge framework for reliable, high-throughput evaluation.

  • WARP joins ICML'26 WSS with a technique that recovers training data mixtures from model weights.

  • I receive an ICML'26 Gold Reviewer recognition. Thank you, ICML!

  • Two papers join ICML'26: Grad-Mimic, for efficient data selection using weight-space geometry, and CARE, for confounder-aware aggregation in LLM evaluation.

  • Our new unified scaling laws work studies how to jointly optimize pretraining and test-time compute.

  • We introduce RubiCap, a rubric-guided RL framework for high-quality image captioning.

  • We launch Data Foundations of AI, a new online seminar to support the data-centric AI community.

  • I will give four talks about Foundation Data Recipes in Taiwan, hosted by TSMC, NTU, NTHU, and NYCU.

  • I return to Apple as a research intern, building synthetic verifiers for open-ended visual RL.

  • WEAVER joins NeurIPS'25 with a framework that aggregates weak verifiers into a powerful one.

  • I will give a talk about automated labeling and data selection at Stanford's Scaling Intelligence Lab.

  • We share early results for PAJAMA, a program distillation approach to automate evaluation cheaply.

  • Grad-Mimic receives an oral presentation at ICML'25 DataWorld workshop. Check out the recording!

  • I join Meta GenAI as a research intern, working on synthetic data generation and off-policy distillation.

  • I pass my Ph.D. preliminary exam. Many thanks again to my graduate committee!

  • I give a talk about spatio-temporal modeling for sensor networks at NTUT in Taiwan.

  • We introduce R&B, a new data mixing method for efficient model training.

  • We present Mimic Score and Grad-Mimic, using model weight geometry to distill training data insights.

  • Alchemist appears at NeurIPS'24 (Spotlight), offering a 500x cheaper way to label data with programs.

  • We release MoRe, a new parameter-efficient fine-tuning method integrated with Monarch matrices.

  • I join Apple as a research intern, focusing on data curation and data mixing for pretraining.

  • I pass my Ph.D. qualification exam. Many thanks to my graduate committee!

  • I receive a NeurIPS'23 Scholar Award. Thank you, NeurIPS!

  • Two papers appear at NeurIPS'23: Train 'n Trade, on building parameter marketplaces for trading model weights, and Loki, about geometry-aware adaptation on model predictions.

  • Our data curation strategy, built with object detection and filter ensembles, ranks first on the DataComp'23 leaderboard and appears at the ICCV'23 TNGCV workshop.

  • We are running the AutoML Cup, a new competition for automated machine learning.

  • ScriptoriumWS appears at ICLR'23 DL4C workshop as a code generation assistant for weak supervision.

  • AutoWS-Bench-101 presents at NeurIPS'22, a new benchmark for comparing weak supervision methods.

  • I serve as president of the Student Association of Taiwan (SAT) at UW-Madison.

  • I receive a First-year Departmental Scholarship from UW-Madison. Thank you, CS Department!

  • I serve as vice president of the Student Association of Taiwan (SAT) at UW-Madison.

  • Here is the start of my Ph.D. journey at UW-Madison. WHAT'S NEXT?!

Experience

  • Apple logo
  • Meta logo
  • University of Wisconsin-Madison logo
  • Argonne National Laboratory logo
  • Academia Sinica logo
  • National Chengchi University logo
Resume [PDF]

Education

  1. University of Wisconsin, Madison (UW-Madison)

    Aug. 2021 — Aug. 2026

    Ph.D. in Computer Science.

  2. National Chengchi University (NCCU)

    Sep. 2016 — Jul. 2020

    B.S. in Computer Science.

Publications and Preprints

  1. Codifying the Judge: Scalable Evaluation via Program Distillation

    Under Submission

    Tzu-Heng Huang*, Shengqi Qiu*, Frederic Sala

    [PDF] [CODE] [PROJECT PAGE]

  2. Test-Time Scaling Makes Overtraining Compute-Optimal

    Under Submission

    Nicholas Roberts, Sungjun Cho, Zhiqi Gao, Tzu-Heng Huang, Albert Wu, Gabriel Orlanski, Avi Trost, Kelly Buchanan, Aws Albarghouthi, Frederic Sala

    [PDF] [X POST]

  3. RubiCap: Rubric-Guided Reinforcement Learning for Dense Image Captioning

    Under Submission

    Tzu-Heng Huang, Sirajul Salekin, Javier Movellan, Frederic Sala, Manjot Bilkhu

    [PDF] [X POST]

  4. WARP: Weight-Space Analysis for Recovering Training Data Portfolios

    ICML'26 Weight-Space Symmetries: from Foundations to Practical Applications (WSS) Workshop

    Tzu-Heng Huang*, Aditya Goyal*, John Cooper, Frederic Sala

    [PDF] [CODE]

  5. Evaluating Sample Utility For Efficient Data Selection by Mimicking Model Weights

    ICML'26 ICML'25 Unifying Data Curation Frameworks Across Domains (DataWorld) Workshop (Oral)

    Tzu-Heng Huang, Manjot Bilkhu, John Cooper, Frederic Sala, Javier Movellan

    [PDF] [CODE] [X POST]

  6. CARE: Confounder-Aware Aggregation for Reliable LLM Evaluation

    ICML'26

    Jitian Zhao, Changho Shin, Tzu-Heng Huang, Satya Sai Srinath Namburi GNVV, Frederic Sala

    [PDF] [CODE] [X POST]

  7. Time to Impeach LLM-as-a-Judge: Programs are the Future of Evaluation

    ICML'25 Programmatic Representations for Agent Learning (PRAL) Workshop

    Tzu-Heng Huang, Harit Vishwakarma, Frederic Sala

    [PDF] [CODE] [X POST]

  8. Shrinking the Generation-Verification Gap by Scaling Compute for Verification

    NeurIPS'25 ICML'25 Efficient Systems for Foundation Models (ES-FoMo III) Workshop ICML'25 Multi-Agent Systems in the Era of Foundation Models: Opportunities, Challenges and Futures (MAS) Workshop

    Jon Saad-Falcon, E. Kelly Buchanan, Mayee F Chen, Tzu-Heng Huang, Brendan McLaughlin, Tanvir Bhathal, Shang Zhu, Ben Athiwaratkun, Frederic Sala, Scott Linderman, Azalia Mirhoseini, Christopher Re

    [PDF] [CODE] [BLOG] [X POST]

  9. From Many Voices to One: A Statistically Principled Aggregation of LLM Judges

    NeurIPS'25 Evaluating the Evolving LLM Lifecycle: Benchmarks, Emergent Abilities, and Scaling (LLM-Eval) Workshop NeurIPS'25 Reliable ML from Unreliable Data (Reliable ML) Workshop

    Jitian Zhao, Changho Shin, Tzu-Heng Huang, Satya Sai Srinath Namburi GNVV, Frederic Sala

    [PDF] [CODE] [X POST]

  10. R&B: Domain Regrouping and Data Mixture Balancing for Efficient Foundation Model Training

    ICML'25 Unifying Data Curation Frameworks Across Domains (DataWorld) Workshop ICML'25 Data in Generative Models (The Bad, the Ugly, and the Greats) (DIG-BUGS) Workshop

    Albert Ge, Tzu-Heng Huang, John Cooper, Avi Trost, Ziyi Chu, Satya Sai Srinath Namburi GNVV, Ziyang Cai, Kendall Park, Nicholas Roberts, Frederic Sala

    [PDF] [X POST]

  11. The ALCHEmist: Automated Labeling 500x CHEaper Than LLM Data Annotators

    NeurIPS'24 (Spotlight)

    Tzu-Heng Huang, Catherine Cao, Vaishnavi Bhargava, Frederic Sala

    [PDF] [CODE] [BLOG] [X POST]

  12. MoRe Fine-Tuning with 10x Fewer Parameters

    ICML'24 Efficient Systems for Foundation Models (ES-FoMo) Workshop ICML'24 Foundation Models in the Wild (FM-Wild) Workshop

    Wenxuan Tan, Nicholas Roberts, Tzu-Heng Huang, Jitian Zhao, John Cooper, Samuel Guo, Chengyu Duan, Frederic Sala

    [PDF] [CODE]

  13. Train 'n Trade: Foundations of Parameter Markets

    NeurIPS'23

    Tzu-Heng Huang, Harit Vishwakarma, Frederic Sala

    [PDF] [X POST]

  14. Geometry-Aware Adaptation for Pretrained Models

    NeurIPS'23

    Nicholas Roberts, Xintong Li, Dyah Adila, Sonia Cromp, Tzu-Heng Huang, Jitian Zhao, Frederic Sala

    [PDF] [CODE] [X POST]

  15. Multimodal Data Curation via Object Detection and Filter Ensembles

    ICCV'23 Towards the Next Generation of Computer Vision Datasets (TNGCV) Workshop 1st place on the Datacomp leaderboard (small-scale filtering track)

    Tzu-Heng Huang*, Changho Shin*, Sui Jiet Tay, Dyah Adila, Frederic Sala

    [PDF] [X POST]

  16. ScriptoriumWS: A Code Generation Assistant for Weak Supervision

    ICLR'23 Deep Learning for Code (DL4C) Workshop 2023 Midwest Machine Learning Symposium

    Tzu-Heng Huang, Catherine Cao, Spencer Schoenberg, Harit Vishwakarma, Nicholas Roberts, Frederic Sala

    [PDF] [CODE]

  17. AutoWS-Bench-101: Benchmarking Automated Weak Supervision with 100 Labels

    NeurIPS'22

    Nicholas Roberts, Xintong Li, Tzu-Heng Huang, Dyah Adila, Spencer Schoenberg, Cheng-Yu Liu, Lauren Pick, Haotian Ma, Aws Albarghouthi, Frederic Sala

    [PDF] [CODE] [BLOG] [X POST]

  18. Key Sensor Discovery for Quality Audit of Air Sensor Networks

    MobiSys'20

    Tzu-Heng Huang, Cheng-Hsien Tsai, Man-Kwan Shan

    [PDF]

Experience

  1. AIML Research Intern, Apple

    Oct. 2025 — Jan. 2026, advised by Javier Movellan and Manjot Bilkhu

  2. Research Intern, Meta GenAI (now MSL)

    May. 2025 — Sep. 2025, advised by Ernie Chang, Sang Michael Xie, Yiting Lu, and David Kant

  3. AIML Research Intern, Apple

    May. 2024 — Dec. 2024, advised by Manjot Bilkhu and Javier Movellan

  4. CEO & Co-founder, Awan.AI LLC

    May. 2023 — Apr. 2024, established with Eric Lin and Jet Lin

  5. Graduate Research Student, University of Wisconsin-Madison

    Feb. 2022 — Present, advised by Frederic Sala

  6. Research Intern, Argonne National Laboratory

    Jun. 2019 — Sep. 2019, advised by Charlie Catlett and Rajesh Sankaran

  7. Research Assistant, National Chengchi University

    Sep. 2018 — Aug. 2021, advised by Man-Kwan Shan

  8. Research Intern, Academia Sinica

    Feb. 2018 — Jul. 2020, advised by Ling-Jyh Chen

  9. Research Assistant, National Chengchi University

    Jul. 2017 — Jul. 2020, advised by Changya Hu