Portrait of Xucheng Yu

Xucheng Yu

Trustworthy AI • LLM Safety and Reliability • Agentic and Multi-Agent Systems

M.Eng. in Electrical & Computer Engineering, University of Illinois Urbana-Champaign

xy63@illinois.edu · +86 187 6288 3800 · Urbana-Champaign, IL

About

I am a researcher working on trustworthy foundation models. My work centers on adversarial robustness, auditable reasoning and evaluation, and reliable agentic and multi-agent decision-making under uncertainty. I am drawn to the questions of how large language models behave when they are pushed, probed, and deployed at scale — and how we can make them safer and more dependable in the process.

I recently completed my Master of Engineering in Electrical and Computer Engineering at the University of Illinois Urbana-Champaign (GPA 4.0/4.0), where I work with Dr. Haohan Wang and Dr. Huan Zhang. Before UIUC, I earned dual B.Sc. degrees in Mathematics with Computer Science from the Guangdong Technion — Israel Institute of Technology (GTIIT) and the Technion — Israel Institute of Technology, graduating on the Dean's List. Alongside research, I build production systems — from GEO auditing platforms to natural-language tool orchestration.

News

Publications & Manuscripts

Author names in bold indicate my contribution. Submitted manuscripts are under peer review.

  1. SCI-Defense: Detecting Persuasion-Structured Manipulation of LLM-Based Ranking in Generative Engines
    Xucheng Yu, Haibo Jin, Huimin Zeng, Haohan Wang
    Submitted to The Web Conference (WWW) 2027 arXiv:2605.21948
    [PDF] [arXiv]
  2. MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models
    Han Wang, Yifan Sun, Brian Ko, Mann Talati, Jiawen Gong, Zimeng Li, Naicheng Yu, Xucheng Yu, Wei Shen, Vedant Jolly, Huan Zhang
    Accepted at COLM 2026 arXiv:2603.28590
    [PDF] [arXiv]
  3. Understanding Content Moderation in Large Language Models through Restricted Books: From Refusal to Warning
    Xucheng Yu, Emily Knox, Haohan Wang
    Accepted at AIES 2026 arXiv:2608.11806
    [PDF] [arXiv]
  4. HEART: Harness Engineering in LLM Tool Use via Agent-Native Reusable Tool Primitives
    Haibo Jin, Suijin Wang, Xucheng Yu, Haojing Luo, Haohan Wang
    Submitted to ICLR 2027 arXiv:2609.01736
    [PDF] [arXiv]
  5. Prompt Stability Matters: Evaluating and Optimizing Auto-Generated Prompt in General-Purpose Systems
    Ke Chen, Xucheng Yu, Yufei Zhou, Haohan Wang
    Accepted at CPAL 2026 arXiv:2505.13546
    [PDF] [arXiv]
  6. Closed-Loop Self-Improving Scientific Paper Generation: An Iterative Framework Combining AI Generation, Automated Review, and Targeted Optimization
    Xucheng Yu, Ruojia Tao, Haohan Wang
    Manuscript
    [PDF]
  7. Gap Detection: Mitigating Homogenization through Demand and Defensibility
    Xucheng Yu, Haibo Jin, Xiaoqi Wang, Runze Kong, Haohan Wang
    Submitted to ICLR 2027 arXiv:2610.20162
    [PDF] [arXiv]
  8. Where Does Rhetorical Distortion Live? A Multi-Platform, Multi-Language Measurement of Influencer Content
    Xucheng Yu, Jiangping Chen, Haohan Wang
    Submitted to The Web Conference (WWW) 2027 arXiv:2610.26412
    [PDF] [arXiv]
  9. GEO Survey and Unified Evaluation Benchmark
    Xiaoqi Wang, Xucheng Yu, Runze Kong, Haibo Jin, Huimin Zeng, Yifan Xu, Haohan Wang
    Submitted to IEEE TKDE
    [PDF]

Research Experience

Research Assistant, University of Illinois Urbana-Champaign

Gap Detection: Mitigating Homogenization through Demand and Defensibility Jul – Oct 2026
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC
  • Formulated a unified Gap Detection framework connecting three forms of AI content homogenization — LLM output diversity, GEO market positioning, and synthetic training-data collapse — via Demand × Defensibility optimization in embedding space.
  • Developed the Unmet-Demand Theorem and a Nash-product objective; reduced LLM response collision by 12.8% on curated topics and 10.8% across hundreds of INFINITY-CHAT responses.
Rhetorical Distortion across Platforms and Languages Jul – Oct 2026
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC · Co-advisor: Dr. Jiangping Chen, Professor, School of Information Sciences, UIUC
  • Measured rhetorical distortion across 13,060 English and Chinese posts from 330 accounts on six platforms (RSS/Newsletter, YouTube, Bluesky, Reddit, Weibo, Twitter/X); found platform more strongly associated with distortion than content category (partial η² 0.49 vs. 0.09), with X/Twitter replies carrying 3.5× the distortion of the tweets they answer.
  • Built a three-tier pipeline combining bilingual lexical rules, negative-pattern filtering, and GPT-4o-mini verification over five distortion dimensions, calling the LLM on only 14% of posts; validated on 198 posts labeled by two human annotators (inter-annotator Cohen's κ = 0.76).
GEO Survey and Unified Evaluation Benchmark May – Oct 2026
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC
  • Systematically characterized adversarial manipulation of LLM ranking systems across 20 GEO pipelines, establishing a benchmark foundation for evaluating trustworthiness under deployment-time attacks.
  • Compared AutoGEO, MAGEO, AgenticGEO, and E-GEO to surface algorithmic assumptions and new optimization and defense directions.
SCI-Defense: Semantic Manipulation Defense for LLM Ranking Jan – May 2026
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC
  • Designed a defense in which an LLM judge scores five persuasion devices, complemented by perplexity and relevance gates; detected 89.6% of reasoning attacks and 96.2% of review attacks at a 1.2% false-positive rate across 1,152 CORE attacks on Amazon and MS MARCO.
  • Blocked 96.1% of attacks that pushed their target into the top three while displacing only 0.6% of honest items, versus 4.0–4.5% for paraphrasing and truncation; mapped failure modes with six black-box attacks, including relevance inflation.
HEART: Reliable Multi-Agent Tool Use Jan – May 2026
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC
  • Designed Tool Primitives and HEART's Planner-Router-Verifier workflow; built ToolFace, a repository of 25,519 functions supporting dynamic retrieval, nested calls, and feedback-driven recovery.
  • Evaluated on five benchmarks with a 10% gain over SFT baselines and 6% over frontier commercial models, while reducing token consumption by up to 85%.
Content Moderation through Restricted Books Jan – May 2026
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC
  • Executed a 40,800-pair empirical study across 400 books and 17 prompt designs over six frontier models (Claude Sonnet 4.5, GPT-4o, Gemini 2.5 Flash, DeepSeek V3, Qwen-Plus, Grok-4.1-Fast); uncovered a near-zero refusal rate (0.07%), revealing systematic cross-model safety-alignment gaps.
  • Characterized the shift from refusal to warning-based moderation, with warning gaps of 8–15 points and prompt-dependent gaps reaching 19 points.
MonitorBench: Chain-of-Thought Monitorability Jan – Apr 2026
Advisor: Dr. Huan Zhang, Assistant Professor of ECE at UIUC
  • Contributed to MonitorBench, a benchmark evaluating chain-of-thought monitorability across 1,514 instances, 19 tasks, and seven categories, with standard, direct-concealment, and monitor-aware-evasion settings.
  • Implemented the impossible-coding-task evaluation adapted from ImpossibleBench and engineered a parallel multi-model inference and verification pipeline for large-scale experiments.
Prompt Stability and Multi-Agent Systems for Data Science Feb – Sep 2025
Advisor: Dr. Haohan Wang, Assistant Professor, School of Information Sciences, UIUC
  • Established semantic stability as a key reliability criterion for auto-generated prompts, with an optimization framework linking prompt consistency to system-level task success.
  • Architected a full-stack multi-agent platform with editable and regenerable agent dialogues, reaching 95% task completion and reducing user intervention by 40%.

Research Systems & Platforms

Generative Engine Optimization Platform — TopCited.ai May 2026 – Present
Research Engineer
  • Developing GEO SaaS features for AI visibility auditing, content optimization, and SCI-Defense-based ranking-manipulation detection across ChatGPT, Gemini, and Perplexity.
Natural-Language Tool Orchestration Platform — Heelo May 2026 – Present
Research Engineer
  • Engineering natural-language tool orchestration with dynamic routing, multi-turn clarification, browser-based integrations, and credential-isolated execution.

Professional Experience

Donovan's Piano Room Feb – Jul 2026
Full-Stack Engineer · Urbana-Champaign, IL
  • Resolved authentication and session-management defects across the Next.js frontend, Node.js/TypeScript backend, and Redis infrastructure by redesigning token-validation logic and decoupling protected routes from short-lived access tokens.
Bosch Automotive Products (Suzhou) Co., Ltd. Sep – Oct 2023
Software Engineering Intern · Suzhou, China
  • Developed and deployed an IoT data-processing and machine-learning pipeline using Airflow, Pandas, TensorFlow, FastAPI, and C/Cython — reducing preprocessing time by 40%, improving model accuracy by 12%, lowering query latency by 30%, and accelerating inference by 10×.

Projects

Augmented Reality Interactive Modeling for Kohler May – Aug 2025
  • Developed a real-time AR inspection pipeline using Python, Open3D, and Unity — achieving 0.1–0.2 mm precision and ICP-based pose tracking for meshes with over 10 million vertices.
RLHF for Qwen2.5-7B-Instruct Jan – May 2025
  • Implemented an end-to-end PPO/GAE/KL-penalty pipeline with LoRA; achieved 70.1% pairwise preference accuracy and a 64.7% win rate versus 33.0% for the baseline on Stack Exchange data.
Distributed Consensus and Fault Detection (Paxos / Raft) Jan – Apr 2025
  • Implemented Paxos and Raft over gRPC/Protobuf with 99.9% leader-election success and sub-200 ms election latency under network partitions; halved fault-detection time in simulated failures.

Teaching

Peer Tutor & Grader
Guangdong Technion — Israel Institute of Technology
  • Tutored lower-year students in mathematics and computer science, supporting concept clarification, problem-solving strategies, and adaptation to a rigorous Technion-aligned curriculum.
  • Graded assignments for Calculus, Topology, and Probability, giving feedback on mathematical reasoning, proof structure, and solution clarity.

Skills

Research Methods

LLM evaluation, benchmark design, adversarial testing, large-scale empirical analysis, RLHF, multi-agent system evaluation

Programming

Python, C/C++, Go, Java, TypeScript/JavaScript

Frameworks & Tools

PyTorch, vLLM, SGLang, LoRA, FastAPI, Docker, React, Node.js, gRPC/Protobuf