InteractEval
Define interactive world modeling and human-centered AI evaluation.
Worldeology Lab / Faculty
Co-Founder & Chief Cognitive Scientist of AlpsLab · Founder of Worldeology
Xuanming
Zhang
Xuanming Zhang is Co-Founder and Chief Cognitive Scientist of AlpsLab. His research program, Cognizer, explores how AI can move beyond closed-ended pattern matching and distribution alignment through a test-time cognition flywheel that infers latent states, organizes chains of logic, recursively examines and revises its reasoning, models long-horizon consequences, transfers learned patterns across domains, and creates structures that do not yet exist.
His work has appeared in NeurIPS, EMNLP, ACL, ICLR, and PNAS, including two Oral/Spotlight awards. He earned his bachelor’s degree from the University of Wisconsin–Madison and was a joint undergraduate researcher at Stanford University. His prior research spans Qwen (Subjectivity & Interactivity Team), Anthropic (Interpretability Team), Amazon (AGI Team), ByteDance (Coze Team), and Zhipu AI (Emohaa Team).
His long-term ambition is to build AI with cognition, agency, and enduring world models—systems capable not only of reasoning, but of sustained innovation and organization in the real world. Outside AI research, he is a professional psychological counselor and has worked with more than 50 clients.
01
Recent news
Confident Decoding
Accepted to EMNLP 2026 in the top 0.45%; released with an official QwenLM implementation.
Cognition-of-Thought
Accepted to NeurIPS 2026, extending metacognitive reasoning into an auditable rollback-and-refine process.
Long-horizon AI behavior
Work on deceptive interaction and covertly misaligned advice appeared at ICLR and in PNAS.
MetaMind
Received a NeurIPS Spotlight for modeling human social thought with metacognitive multi-agent systems.
02
Research focus
Cognition is not a single answer. It is a trajectory that begins with another mind, passes through a revisable act of reasoning, and continues into the world that action changes.
Xuanming’s work connects model mechanisms, social cognition, and interactive world modeling. The goal is to make latent cognitive states observable, reasoning processes correctable, and downstream consequences measurable—then use those signals to train more capable and trustworthy systems.
- 01
Latent human states
Inferring beliefs, intentions, preferences, emotions, and awareness before a system acts.
- 02
Reasoning boundaries
Making reasoning inspectable, revisable, and able to improve its own cognitive process.
- 03
Interactive worlds
Evaluating how actions alter people, relationships, and long-horizon trajectories.
- 04
Cognizer systems
Moving AI from closed-ended response optimization toward innovation, organization, and creation.
03
Selected publications
Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding
Xuanming Zhang, Sining Zhoubian, Yuxuan Chen, Tianyi Tang, An Yang, et al.
Cognition-of-Thought Elicits Social-Aligned Reasoning in Large Language Models
Xuanming Zhang, Yuxuan Chen, Samuel Yeh
Simulating and Understanding Deceptive Behaviors in Long-Horizon Interactions
Yang Xu*, Xuanming Zhang*, Samuel Yeh, Jwala Dhamala, Ousmane Dia, Rahul Gupta
Human Preferences Are Susceptible to Covertly Misaligned AI Advice
Sahand Sabour, June M. Liu, Siyang Liu, Chris Z. Yao, Shiyao Cui, Xuanming Zhang, et al.
MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
Xuanming Zhang, Yuxuan Chen, Samuel Yeh
SocialEval: Evaluating Social Intelligence of Large Language Models
Jinfeng Zhou, Yuxuan Chen, Yihan Shi, Xuanming Zhang, et al.
04
Experience & education
Research & industry
- 2026—
- WorldeologyFounder
- 2026—
- AlpsLabCo-Founder & Chief Cognitive Scientist
- 2026
- Qwen, Alibaba GroupResearcher
- 2026
- Anthropic (Interpretability Team)Research Fellow
- 2025
- Stanford NLP GroupVisiting Undergraduate Researcher
- 2025
- Amazon (AGI Team)Student Researcher · long-horizon AI safety
- 2024—25
- ByteDance (Coze Team) · Zhipu AI (Emohaa Team)AI research, architecture and agent systems
Education
- 2025—26
- Stanford UniversityJoint Undergraduate in Computer Science · GPA 4.21/4.00
- 2024—26
- University of Wisconsin–MadisonBachelor of Science in Information Science
Service & recognition
NeurIPS Top Reviewer 2025; program committee service for ICLR, ACL, OOPSLA, AAAI, TMLR, and IEEE TII. Invited talks include WAIC, Fudan University, and ByteDance.