Research
My research centers on rethinking the primitive building blocks of deep learning architectures, and how new designs change what a model can learn and at what cost. Humans can learn a new card game from a few examples, revise a plan mid-execution, and transfer knowledge across domains; today's models, despite their scale, largely cannot. I'm drawn to the architectural mechanisms that might close that gap, rather than to scaling existing designs further.
My first-author work, PolyNeXt (ICML 2026), is one instance of this: it replaces standard nonlinearities (ReLU, GELU, softmax) with activation-free polynomial, Hadamard-product modules that match or exceed activation-based vision backbones at lower compute. One of its lessons is that depth, not width, drives expressiveness: stacking narrow layers buys exponentially richer functions at only linear parameter cost. I'm extending this architecture-first program toward new attention mechanisms and layer primitives, redesigning the internals of how large models are built rather than treating them as fixed.
That depth-over-width lesson also points past simply stacking deeper networks. More broadly, I'm interested in alternative architectural designs that let a model reason and deliberate in a learned latent space, using recurrence to add reasoning depth rather than by stacking more layers. World models are one example of this, learning structured internal representations a model can predict and plan over.
Publications
-
Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models
ICML 2026 project page arXiv code -
Challenges and Future Directions of Data-Centric AI Alignment
ICML 2025 (Position Paper) arXiv -
Polynomial Vision Networks: Architecture and the Limits of Block-Recurrent Compression
B.S. Honors Thesis in Computer Science, University of Wisconsin–Madison, 2026 thesis
Experience
Researcher, Chrysos Lab, UW-Madison
First-author work on PolyNeXt (ICML 2026): a family of CNN- and transformer-style backbones (PolyMLP, PolyConv, PolyAttn) built from activation-free polynomial (Hadamard-product) units in place of ReLU and softmax. The models reach 84.3% top-1 on ImageNet-1K at 26M parameters and scale to 85.2%, matching the activation-based state of the art (CAFormer) at equal size while staying fully activation-free, and surpassing prior polynomial networks by ~2-3% at less than half the FLOPs and 30-45% less peak GPU memory than comparable MetaFormer models, with strong out-of-distribution robustness (ImageNet-C/A/R/Sketch) and ADE20K segmentation transfer. A fully-polynomial (BatchNorm) variant uses only adds and multiplies, a step toward FHE-compatible inference.
My senior honors thesis extends this line, showing empirically that block-recurrent depth compression, which works well on trained ViTs, fails on polynomial networks: their depth computes genuinely distinct polynomial terms rather than redundant refinements.
project page · arXiv · code
Researcher, Sharon Li Lab, UW-Madison
Co-authored an ICML 2025 position paper on data-centric AI alignment, arguing that current methods (RLHF, DPO) overweight algorithms relative to data quality. Ran experiments quantifying how noise in preference data degrades reward models, and designed a noise-robust preference-learning objective (Generalized Binary Cross-Entropy) that improved training stability and accuracy over label-smoothing baselines.
Co-founder & Machine Learning Engineer, Correkt AI
Built an LLM-powered multimodal search system serving 40K+ users, with funding and grants from CNSI, the Crescent Fund, and SBAA. Fine-tuned LLaMA-3-70B with RLHF, designed synthetic-data pipelines to generate and rank millions of preference pairs at <1% the cost of human annotation, and implemented web-scale retrieval over 800M+ indexed pages with ~300ms query latency.
Software Development Engineer Intern (Return), Amazon Web Services
Building a recommendation system that selects optimal inclusion-key sets to de-duplicate CloudTrail events for IAM Access Analyzer, designed to compress ~100 TB/hr of event data by 98%+. Engineering the distributed pipeline with Step Functions, Lambda, and a PySpark Glue ETL over Parquet.
Software Development Engineer Intern, Amazon
Built a high-throughput CDC pipeline (~100K events/s) from DynamoDB to Apache Iceberg, enabling real-time ML feature stores and analytics. Delivered a reusable AWS CDK construct that cut data integration time from weeks to minutes.
Software Engineer Intern, 2Sigma School
Fine-tuned LLM agents and NLP pipelines for personalized feedback and automatic question generation in CS education, using in-context learning and code-execution-based hallucination detection. Cut teacher prep time by 30% and raised satisfaction by 23%, helping deliver personalized curricula to thousands of students.
Projects
Evaluating Modality Bias in Multimodal LLMs
Built a counterfactual-ablation pipeline to measure how multimodal LLMs (Qwen2.5-Omni-7B, Gemma-3n) weight visual versus textual evidence, with custom metrics (Attribution Faithfulness Score, Lexical Override Rate, Text Override Rate). Models misreport their own modality use and flip 69% of correct answers under adversarial text injection; modality-separated prompting and importance scoring did not reliably beat baselines, suggesting fixes require architecture or training changes rather than prompting.
Depth Efficiency of Recurrent Networks
Extended depth-efficiency theory (why deeper networks are exponentially more expressive) to recurrent models, proving via measure theory that a Tensor-Train network is RNN-equivalent with the same exponential rank separation.
Quantum-Inspired SQL Query Optimization
Modeled database join ordering as a QUBO problem and built a quantum join-reordering pipeline (QAOA, VQE) in Qiskit, with SQL-to-QUBO compilation from relational query plans. Achieved 15% lower average estimated cost vs. the PostgreSQL baseline plan on 5+ table TPC-H queries.
Honors
- Distinguished Graduate Fellowship, University of Wisconsin-Madison (2026)
- Dean's List (Fall 2022 – Spring 2025), University of Wisconsin-Madison; GPA 4.0/4.0
- USACO Gold (2021)
Service
- Reviewer, ICLR 2026 Workshop on Agentic AI in the Wild: From Hallucinations to Reliable Autonomy
About
I'm finishing a B.S. in Computer Science (Honors) and Data Science at UW-Madison and beginning my Ph.D. in Electrical & Computer Engineering (ML focus) there in Fall 2026, advised by Grigorios Chrysos and supported by a Distinguished Graduate Fellowship. Outside of research, I enjoy puzzle games like Baba Is You and grand strategy games like Crusader Kings.
I'm always glad to hear about research opportunities or interesting problems to work on together, whether it's a role, a grant, or just an idea worth exploring. The best way to reach me is by email.