Hello! This is Haizhao Yang.

Recent Works

Mathematical foundations of deep learning, symbolic and interpretable ML, agentic AI for computational science, and scientific foundation models

Agon system overview
Agentic AI

Agon: An Autonomous Large-Scale Omnidisciplinary Research System Built on Prompt Economy

Youran Sun, Xingyu Ren, Chugang Yi, Jiaxuan Guo, Kejia Zhang, Jianda Du, Haizhao Yang

The world has entered the age of agents, yet prompt engineering is still the Wild West, with no shared rules of thumb. We distill six design principles for autonomous research, anchored by Prompt Economy: stop writing prompts for single tasks, and start building reusable loops. On this foundation we built Agon, an omnidisciplinary AI Scientist whose deployments reveal many intriguing phenomena, including agent emergence, and we map recurring failure modes into a boundary map. Our dream: human navigates, AI drives.

arXiv PDF Code
→
FEX TranNet prediction map
Scientific AI

Finite Expression Method with TranNet-based Function Learning for High-Dimensional PDEs

Toan Huynh, Feng Bao, Haizhao Yang, Ahmed Zytoon

This project turns intimidating PDEs into a formula hunt that feels almost cinematic. FEX searches through expression space while TranNet supplies learned mathematical operators, giving the solver a richer bag of tricks than hand written symbols alone. The payoff is bold and visual: huge equations become compact shapes people can inspect, reuse, and understand.

arXiv PDF
→
DRIFT hallucination routing pipeline
LLM Reliability

DRIFT: Detecting Representational Inconsistencies for Factual Truthfulness

Rohan Bhatnagar, Youran Sun, Chi Zhang, Yixin Wen, Haizhao Yang

DRIFT moves hallucination detection inside the model before the answer finishes. It reads risk signals from intermediate hidden states with less than 0.1% extra compute and runs in parallel with generation. The main result is clear: SOTA AUROC on 10 of 12 settings across four QA benchmarks and three model families, with gains up to 13 points.

arXiv PDF ResearchGate DBLP SciRate
→
GA-ICL dialogue hallucination detection results
LLM Reliability

Geometry-Aware Hallucination Detection in Large Language Models

Bodla Krishna Vamshi, Rohan Bhatnagar, Haizhao Yang

GA-ICL turns in-context example selection into a geometry problem. It learns local manifolds and class prototypes inside frozen LLM representation space, then chooses demonstrations that better separate factual consistency. On FEVER and HaluEval it beats standard ICL retrieval baselines in most settings, with especially strong gains on dialogue and summarization.

arXiv PDF SciRate
→