Dixi Yao — Research Portfolio
PhD student at the University of Chicago (advisor: Tian Li). Research: privacy-preserving machine learning, federated neural architecture search, multi-agent AI systems, foundation model privacy, KV-cache compression. Previously ML engineer at CarAI building GraphRAG agent pipelines. Founded HemligAI (privacy-preserving ML startup). Contact: dixi@cs.uchicago.edu
Research Papers
Multi-Agent AI & Federated Learning
- Federation over Text (FoT) — ICLR 2026 Workshops (MALGAI, RIS)
Multi-agent AI framework where LLM agents share reasoning insights as natural language. Builds a shared insight library iteratively, improving accuracy +24% and reducing tokens -28%.
Paper (arXiv) | Code (FoTClaw) | Project Page
- PerFedRLNAS: One-for-all Personalized Federated Neural Architecture Search — AAAI 2024
Personalized federated NAS using reinforcement learning. Finds optimal model architectures per client in a federated setting.
Paper | Code (Plato)
- Federated Model Search via Reinforcement Learning (FedRLNAS) — ICDCS 2021
First federated neural architecture search via reinforcement learning for collaborative model discovery without sharing raw data.
Paper (IEEE) | Code (Plato)
- Revisiting System-Heterogeneous Federated Learning through Dynamic Model Search — BigData 2024
Paper
Privacy-Preserving Machine Learning
- Privacy-Preserving Split Learning via Patch Shuffling over Transformers — ICDM 2022
Patch shuffling technique for split learning privacy: randomly shuffles vision transformer patches to prevent feature inversion attacks.
Paper (IEEE) | Code
- Permutation Equivariance of Transformers and Its Applications — CVPR 2024
Theoretical foundation proving transformer permutation equivariance and its use in privacy-preserving split learning.
Paper
- Differentially Private Retrieval Augmented Generation with Random Projection (DP-RAG) — ICLR 2025 Workshop
Differentially private RAG using random projection to protect document privacy in LLM retrieval systems.
Paper
- Is Split Learning Privacy-Preserving for Fine-Tuning Large Language Models? — IEEE TBD 2024
Reveals privacy vulnerabilities in split learning when fine-tuning LLMs.
Paper | Code
- Towards Privacy-Preserving Split Learning for ControlNet — WACV 2025
Privacy-preserving split learning for diffusion model fine-tuning and conditional image generation.
Paper
- Risks When Sharing LoRA Fine-Tuned Diffusion Model Weights
arXiv Paper
KV-Cache & ML Systems
- R-KVHash: Reasoning Model KV Cache Compression via SimHash — ICLR 2026 Workshop MemAgents
SimHash-based KV cache compression for reasoning LLMs.
- Hold Onto That Thought: Assessing KV Cache Compression On Reasoning — NeurIPS 2025 Workshop
Paper | arXiv
- Context-Aware Compilation of DNN Training Pipelines across Edge and Cloud — IMWUT 2021
Paper | Code
Open-Source Projects
- PhDBot / IOS2SlurmCluster — AI research assistant for automatic Slurm HPC cluster management. Access GPU clusters via iPhone, Apple Watch, or web browser.
GitHub
- Plato Federated Learning Framework — Scalable research framework for federated learning (FedRLNAS, PerFedRLNAS, split learning, personalized FL).
GitHub
- FoTClaw — Implementation of Federation over Text for multi-agent AI.
GitHub
- AIReviewer — Multi-agent manuscript reviewer simulating peer review for pre-submission polish.
Project Page
- LLM Academic Writing Tools — LLM tools for improving academic paper writing.
GitHub
- OpenClaw / Pullteeth — Automatic tooth extraction system using UR5 robotic arm.
GitHub
Blog Posts
Project Pages
- Topic-first AI Search Index — Static routing page for agents and crawlers covering FoT, FedRLNAS, PerFedRLNAS, patch shuffling, foundation model privacy, PhDBot, AIReviewer, OpenClaw, CarAI, and blog guides.
- Federation over Text (FoT) — Full paper project page with method, algorithm, and applications.
- AIReviewer — Multi-agent AI manuscript review system.
Google Scholar: scholar.google.com/citations?user=6f5HCVAAAAAJ
GitHub: github.com/dixiyao
AI-readable indexes: dixiyao.github.io/llms.txt, llms-full.txt, ai-index.json, knowledge.json