About
I am Shuhao (Sullivan) Zhang, an M.S. student in the Department of
Computer Science and Engineering at the
University of California, San Diego,
advised by Prof. Pengtao Xie.
Previously, I received my B.S. in Data Science from the
University of Science and Technology Beijing,
where I graduated in the top 7% of my class.
My current research focuses on pre-hoc reasoning and
metacognitive control for LLM systems — using lightweight
estimators and self-monitoring signals to route, defend,
and self-improve LLM agents at test time. I am actively seeking
Research Assistant / Visiting Researcher opportunities in agentic AI,
test-time scaling, and multimodal reasoning.
Selected Publications
LLMs Know When They Know, but Do Not Act on It: A Metacognitive Harness for Test-time Scaling
Qi Cao, Yufan Wang, Peijia Qin, Shuhao Zhang, Pengtao Xie
NeurIPS 2026 Under Review
A metacognitive harness that separates monitoring from reasoning: the model emits
a pre-solve feeling-of-knowing and a post-solve judgment-of-learning signal, turned
into an explicit test-time control interface for retry / aggregation. On a fixed
Claude Sonnet-4.6 base, raised pooled accuracy from 48.3 to 56.9
without any parameter updates, exceeding the strongest leaderboard entries on
HLE-Verified, LiveCodeBench v6, and R-Bench-V.
arXiv ↗
Send a SCOUT First: Pre-hoc Reasoning for Adaptive Detector Allocation in Prompt-Injection Defense
Shuhao Zhang†, Jiarui Li†, Qi Cao, Ruiyi Zhang, Pengtao Xie
EMNLP 2026 Under Review
Reframed prompt-injection defense as per-input detector allocation.
SCOUT summarizes each detector's behavior as a fingerprint over a fixed anchor set,
then trains a small predictor (Qwen3-4B; SFT + GRPO) to estimate, for every new
request, which detectors are reliable and how long each will take. A single operator
threshold trades safety vs. latency. Cuts attack-success rate by 46%
and wall-clock by 40% vs. an always-on GPT-4o judge; transfers
to BIPIA, IPI, and IHEval with no retraining.
arXiv ↗
Models Under SCOPE: Scalable and Controllable Routing via Pre-hoc Reasoning
Qi Cao†, Shuhao Zhang†, Ruizhe Zhou, Ruiyi Zhang, Peijia Qin, Pengtao Xie
ICML 2026 Accepted
SCOPE is a budget-aware LLM routing framework that predicts how accurate and how
expensive each model will be before running it, enabling controllable cost–accuracy
trade-offs and naturally handling new models. Up to 95.1% cost reduction
or 25.7% accuracy gain under different priorities.
Project ↗
IDC-CDR: Cross-domain Recommendation based on Intent Disentanglement and Contrast Learning
Jing Xu, Mingxin Gan, Hang Zhang, Shuhao Zhang
Information Processing and Management, 2024
A cross-domain recommendation model that disentangles user intent for stronger
transfer, with an emphasis on interpretability of latent intent factors.
Using Machine Learning Models for Short-Term Prediction of Dissolved Oxygen in a Microtidal Estuary
Mina Gachloo, Qianqian Liu, Yang Song, Guozhi Wang, Shuhao Zhang, Nathan Hall
Water 16(14):1998, 2024
Applied ML models for short-term prediction of dissolved oxygen, with relevance
to healthcare / biological time-series modeling.
