Shu Yang 杨澍

About me

I am a Ph.D. student in Computer Science at King Abdullah University of Science and Technology (KAUST), advised by Prof. Di Wang in the PRADA Lab (Provable Responsible AI and Data Analytics Lab). I received my M.Sc. in Computational Linguistics from the University of Macau, advised by Prof. Derek F. Wong. I have been a visiting scholar at the University of Edinburgh and Tsinghua University. I am currently a Visiting Student Researcher at the Stanford Digital Economy Lab, Stanford HAI, where I work closely with Sandy Pentland and Jiaxin Pei.

We are building the next generation of human–agent collaboration. Today's AI assistants are private, single-user tools, while real work happens in shared conversations. With Kordi, an open-source, AI-native collaboration workspace, we treat agents as participants in the conversation: people and their agents share the same space, and anyone can call on any agent right where the context lives. This builds on my earlier work on LLM alignment and monitorability, including sycophancy, reward hacking, and reasoning-induced misalignment.

News

• 2026.09 Hearing is Believing? Evaluating and Analyzing Audio Language Model Sycophancy with SYAUDIO accepted by NeurIPS 2026.
• 2026.08 Four papers accepted by EMNLP 2026:
  • LLMs Regret Before They Say It: Early Detection and Compositional Architecture of Regret in Hidden States
  • Understanding and Mitigating Cross-lingual Privacy Leakage via Language-specific and Universal Privacy Neurons
  • Neuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language Models
  • STARS: Skill Triage via Activation Risk Scoring for LLM Agents
• 2026.07 Can AI Truly Represent Your Voice in Deliberations? A Comprehensive Study of Large-Scale Opinion Aggregation with LLMs accepted by CoLM 2026.
• 2026.05 Joined the Stanford Digital Economy Lab at Stanford HAI as a Visiting Student Researcher.
• 2026.04 Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Models accepted by ICML 2026.
• 2026.04 Five papers accepted by ACL 2026:
  • AutoMonitor-Bench: Evaluating the Reliability of LLM-Based Misbehavior Monitor
  • Flattery in Motion: Benchmarking and Analyzing Sycophancy in Video-LLMs
  • Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images Oral
  • Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models
  • JARVIS or Ultron? A Survey on the Safety and Security Threats of Computer-Using Agents
• 2026.02 My first blog, Misalignments and RL failure modes in the early stage of superintelligence, accepted by the ICLR 2026 Blogpost Track.
• 2026.01 Four papers accepted by ICLR 2026:
  • When Thinking Backfires: Mechanistic Insights into Reason-induced Misalignment
  • Neuron-Aware Data Selection in Instruction Tuning for Large Language Models
  • Evaluating Data Influence in Meta Learning
  • Dissecting Representation Misalignment in Contrastive Learning via Influence Function
• 2026.01 Understanding Aha Moments: from External Observations to Internal Mechanisms accepted by TACL.
• 2025.12 The Web Tool Trap: Understanding and Mitigating Over-Reliance in LLM Browsing Agents accepted as an extended abstract by AAMAS 2026.
• 2025.11 When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models accepted by AAAI 2026.
• 2025.10 Our workshop proposal for the First Workshop on LLM Persona Modeling was accepted by NeurIPS 2025. See you in Mexico City!
• 2025.09 Began my visit at EdinburghNLP, University of Edinburgh.
• 2025.09 EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification accepted by NeurIPS 2025.
• 2025.08 Two papers accepted by EMNLP 2025:
  • Can Large Language Models Identify Implicit Suicidal Ideation? An Empirical Evaluation
  • Understanding How Value Neurons Shape the Generation of Specified Values in LLMs
• 2025.08 RepreGuard: Detecting LLM-Generated Text by Revealing Hidden Representation Patterns accepted by TACL.
• 2025.05 Three papers accepted by ACL 2025:
  • Fraud-R1: A Multi-Round Benchmark for Assessing the Robustness of LLM Against Augmented Fraud and Phishing Inducements
  • Understanding the Repeat Curse in Large Language Models from a Feature Perspective (code)
  • Rethinking Prompt-based Debiasing in Large Language Models
• 2025.04 Started my visit at the THUNLP group, Tsinghua University.
• 2024.11 Who Wrote This? The Key to Zero-Shot LLM-Generated Text Detection Is GECScore accepted by COLING 2025.
• 2024.11 Our survey A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions accepted by Computational Linguistics.
• 2024.10 DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios accepted by the NeurIPS 2024 Datasets & Benchmarks Track.
• 2024.08 Began my Ph.D. journey at the PRADA Lab, KAUST, under the supervision of Prof. Di Wang.
• 2024.07 Model Autophagy Analysis to Explicate Self-consumption within Human-AI Interactions accepted by CoLM 2024.
• 2024.06 Completed my Master's degree at the University of Macau, under the supervision of Prof. Derek Wong.

Publications

Selected conference and journal papers. See Google Scholar for the full list.

2026

• NeurIPS 2026 Hearing is Believing? Evaluating and Analyzing Audio Language Model Sycophancy with SYAUDIO
  arXiv
• EMNLP 2026 LLMs Regret Before They Say It: Early Detection and Compositional Architecture of Regret in Hidden States
• EMNLP 2026 Understanding and Mitigating Cross-lingual Privacy Leakage via Language-specific and Universal Privacy Neurons
  arXiv
• EMNLP 2026 Neuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language Models
  arXiv
• EMNLP 2026 STARS: Skill Triage via Activation Risk Scoring for LLM Agents
• CoLM 2026 Can AI Truly Represent Your Voice in Deliberations? A Comprehensive Study of Large-Scale Opinion Aggregation with LLMs
  arXiv
• ICML 2026 Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Models
  arXiv
• ACL 2026 AutoMonitor-Bench: Evaluating the Reliability of LLM-Based Misbehavior Monitor
  arXiv
• ACL 2026 Flattery in Motion: Benchmarking and Analyzing Sycophancy in Video-LLMs
  arXiv
• ACL 2026 Oral Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images
  arXiv
• ACL 2026 Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models
  arXiv
• ACL 2026 JARVIS or Ultron? A Survey on the Safety and Security Threats of Computer-Using Agents
  arXiv
• ICLR 2026 When Thinking Backfires: Mechanistic Insights into Reason-induced Misalignment
  arXiv
• ICLR 2026 Neuron-Aware Data Selection in Instruction Tuning for Large Language Models
  arXiv
• ICLR 2026 Evaluating Data Influence in Meta Learning
  arXiv
• ICLR 2026 Dissecting Representation Misalignment in Contrastive Learning via Influence Function
  arXiv
• ICLR 2026 Blogpost Misalignments and RL failure modes in the early stage of superintelligence
  blogpost LessWrong
• TACL Understanding Aha Moments: from External Observations to Internal Mechanisms
  arXiv
• AAMAS 2026 The Web Tool Trap: Understanding and Mitigating Over-Reliance in LLM Browsing Agents (extended abstract)
• AAAI 2026 When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models
  arXiv

2025

• NeurIPS 2025 EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification
  arXiv
• EMNLP 2025 Can Large Language Models Identify Implicit Suicidal Ideation? An Empirical Evaluation
  arXiv
• EMNLP 2025 Understanding How Value Neurons Shape the Generation of Specified Values in LLMs
  arXiv
• TACL RepreGuard: Detecting LLM-Generated Text by Revealing Hidden Representation Patterns
  arXiv
• ACL 2025 Fraud-R1: A Multi-Round Benchmark for Assessing the Robustness of LLM Against Augmented Fraud and Phishing Inducements
  arXiv
• ACL 2025 Understanding the Repeat Curse in Large Language Models from a Feature Perspective
  arXiv code
• ACL 2025 Rethinking Prompt-based Debiasing in Large Language Models
  arXiv
• COLING 2025 Who Wrote This? The Key to Zero-Shot LLM-Generated Text Detection Is GECScore
  arXiv

2024

• Computational Linguistics A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions
  arXiv
• NeurIPS 2024 D&B DetectRL: Benchmarking LLM-Generated Text Detection in Real-World Scenarios
  arXiv
• CoLM 2024 Model Autophagy Analysis to Explicate Self-consumption within Human-AI Interactions
  arXiv OpenReview

Blog

• A Simple Guide and Best Practices for Using OpenClaw in Research2026/03
• Two New Concepts for Frontier AI Systems: “Context Engineering” and “Chain of Command”2025/10
• Misalignments and RL failure modes in the early stage of superintelligence2025/07ICLR Blogpost Track 2026 · also on LessWrong

Services

• Area Chair: ACL Rolling Review (ARR)
• Conference reviewer: NeurIPS 2024, 2025; ICML 2025, 2026; ICLR 2025, 2026; AISTATS 2025, 2026; ARR (ACL, EMNLP, EACL); AAAI 2026; CVPR 2026; CoLM 2026
• Journal reviewer: Applied Artificial Intelligence; TALLIP; Transactions on Social Computing
• Workshop organizer: First Workshop on LLM Persona Modeling, NeurIPS 2025
• Teaching assistant: KAUST CS 325 Private Data Analysis, Fall 2025

Talks

• 2025.12 Who Is Controlling the Game? Opportunities and Challenges in Bidirectional Human–AI Alignment · EdinburghNLP, University of Edinburgh

Visiting and collaboration

• 2026.05 Visiting Student Researcher, Stanford Digital Economy Lab, Stanford HAI
• 2025.09 Visiting scholar, EdinburghNLP, University of Edinburgh
• 2025.04 Visiting scholar, THUNLP, Tsinghua University