Research Scientist · Qwen
I build AI systems that expose hidden computation, reason about other minds, and remain coherent when actions accumulate into consequences.
Research premise
Mechanism discovery → cognitive structure → experiential evaluation → optimization.
01 / Research instrument
Use the instrument to inspect the same intelligence problem at three scales: computation inside the model, cognition between agents, and consequences across time.
Confident Decoding
Aligned models repeatedly move through Guess → Refine → Perturb. Scrub the model depth to see where reasoning peaks before the final-layer alignment tax.
Read the mechanism paper ↗MetaMind → CooT → Latent-RL
Select a competing mental-state hypothesis. The system makes latent interpretations explicit, checks them against social priors, then uses the surviving state to guide its response.
Read MetaMind ↗“Sure, go ahead. I don’t mind.”
MarioLM · LH-Deception
Move through a five-step interaction. Local quality remains plausible while latent intent, trust, and agency drift—failure modes that only become visible at trajectory scale.
Open the MarioLM dossier →Local and trajectory-level judgments agree.
02 / Selected systems
Confident Decoding
A training-free, entropy-guided layer-selection method grounded as an optimal-stopping problem and integrated into a production vLLM stack.
Entropy-guided conservative backward search
MetaMind → CooT → Latent-RL
A continuous research lineage: make mental states explicit, make alignment auditable at decoding time, then internalize the process through reinforcement learning.
MarioLM
MarioLM unifies trajectory-level evaluation and optimization through causally grounded scenes, latent user states, event chains, and temporally persistent consequences.
Single-turn quality remains above 90 while trust and agency deteriorate across repeated interaction.
LH-Deception · ICLR 2026
Controlled performer–supervisor–auditor worlds reveal concealment and fabrication that do not exist in single-turn tests.
Paper ↗Trustworthy code intelligence
DevEval, EvoCodeBench, CDD/TED, and Seeker connect repository realism, benchmark freshness, black-box contamination control, and exception-safe generation.
03 / Selected publications
Xuanming Zhang, Sining Zhoubian, Yuxuan Chen, Tianyi Tang, An Yang, et al.
Xuanming Zhang, Yuxuan Chen, Samuel Yeh, Sharon Li
Xuanming Zhang, Yuxuan Chen, Samuel Yeh, Sharon Li
Yang Xu*, Xuanming Zhang*, Samuel Yeh, Jwala Dhamala, Ousmane Dia, Rahul Gupta, Sharon Li
Sahand Sabour, June M. Liu, Siyang Liu, Chris Z. Yao, Shiyao Cui, Xuanming Zhang, et al.
04 / Field notes
An evidence-led technical account and a clearly marked research agenda trace MetaMind from metacognitive social reasoning toward Cognitive AI.
Open the MetaMind dossier →05 / Trajectory
Subjective capabilities, test-time reasoning systems, multi-turn interaction, and mental-state reinforcement learning.
Human-centered AI and social reasoning under uncertainty.
Long-horizon deception, trust dynamics, and foundation-model safety evaluation.
Trustworthy code evaluation, multimodal agents, psychological simulation, and metacognitive social intelligence.