Zhou Ziheng

josephziheng [at] ucla [dot] edu

UCLA CS Department

I am a fifth-year PhD candidate in Computer Science at UCLA, co-advised by Demetri Terzopoulos, Song-Chun Zhu, and Yingnian Wu. My research lies at the intersection of artificial intelligence, morality, and consciousness.

Before starting my PhD, I was a co-founding partner at VoxelCloud Inc. (2016–2021), an AI medical imaging company that raised nearly $100M in funding from Tencent, Sequoia China, and others. Our work was published in venues including Nature Medicine and Nature Biomedical Engineering.

I received my B.S. in Computer Science from UCLA (2012–2017), with double minors in Cognitive Science and Philosophy.

My PhD research centers on understanding the nature of three fundamental aspects of mind – intelligence, morality, and consciousness – through formal theory and computational formulation, and on developing AI algorithms and frameworks grounded in these insights (see my academic works). Aside from this framework of designing research questions, my interests can also be described by the broad goal of self-evolving AI (how to be like a scientist that discovers and learns, or like a philosopher that introspects and improves) and human-centered AI (what is good for humans, and how to make it good for humans).

I am actively looking for collaborators at all levels (undergraduate, master, and junior PhD). What I value most is agency and persistency – all major backgrounds are welcome. Feel free to reach out for a chat about potential projects and collaborations.

I am currently on the job market. If you have AI positions or internship opportunities, I would love to hear from you – please feel free to reach out.

news

Apr 2026	Paper accepted at ACL 2026 Findings: “Simple Role Assignment is Extraordinarily Effective for Safety Alignment.”
Apr 2026	Paper accepted at ACL 2026 main conference (Oral): “Why Are We Moral? An LLM-based Agent Simulation Approach to Study Moral Evolution.”
Mar 2026	Paper accepted at ICLR 2026 Workshop RSI (Spotlight): “SciCrafter: Can Current Language Models Close the Discovery-to-Application Loop?”
Mar 2026	Paper accepted at ICLR 2026 Workshop AFAA: “Learning to Be Fair: Modeling Fairness Dynamics by Simulating Moral-Based Multi-Agent Resource Allocation.”
Mar 2026	Paper accepted at ICLR 2026 Workshop AIMS: “Credibility Governance: A Social Mechanism for Collective Self-Correction under Weak Truth Signals.”

selected publications

ACL Oral

Why Are We Moral? An LLM-based Agent Simulation Approach to Study Moral Evolution

Zhou Ziheng, Huacong Tang, Mingjie Bi, and 7 more authors

Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2026

🔥 Oral Abs arXiv Website

true

LLM-based agent simulation framework modeling prehistoric hunter-gatherer societies to investigate moral evolution. Accepted as ACL 2026 main conference oral.
Preprint

Less is More: Early Stopping Rollout for On-Policy Distillation

Zhou Ziheng, Jiaqi Li, Huacong Tang, and 2 more authors

arXiv preprint arXiv:2605.27028, 2026

Abs arXiv

Identifies an Off-policy Teacher Decay problem in on-policy distillation, where teacher models lose effectiveness on later tokens once the student’s earlier trajectory diverges from training data. Proposes Early Stopping Rollout (ESR), which restricts generation to initial response tokens, outperforming full-rollout approaches across model sizes and tasks while improving computational efficiency. Analyzes underlying mechanisms including Cascading Alignment and Sub-mode Commitment.
ICLR WS Spotlight

SciCrafter: Can Current Language Models Close the Discovery-to-Application Loop?

Zhou Ziheng, Huacong Tang, Jinyuan Zhang, and 9 more authors

ICLR 2026 Workshop RSI / Under Review for COLM 2026, 2026

🔥 Spotlight Abs Website

true

We introduce SciCrafter, a Minecraft-based task suite that operationalizes the discovery-to-application loop through parameterized redstone circuit tasks. Evaluating frontier models including GPT-5.2, Gemini-3-Pro, and Claude-Opus-4.5, we find that all plateau at approximately 25% success rate.
ICLR WS

Credibility Governance: A Social Mechanism for Collective Self-Correction under Weak Truth Signals

Wanying He, Yanxi Lin, Zhou Ziheng, and 5 more authors

ICLR 2026 Workshop AIMS, 2026

Abs arXiv

Introduces Credibility Governance (CG), a mechanism that reallocates influence by identifying agents and viewpoints that consistently align with evolving evidence, outperforming vote-based and stake-weighted approaches across misinformation, noise, and adversarial pressure scenarios.
ACL Findings

Simple Role Assignment is Extraordinarily Effective for Safety Alignment

Zhou Ziheng, Jiakun Ding, Zhaowei Zhang, and 6 more authors

Findings of the Association for Computational Linguistics (ACL Findings), 2026

Abs arXiv Website

Proposes role conditioning grounded in Theory of Mind as a compact alternative to principle-based alignment. Reduces unsafe outputs on the WildJailbreak benchmark from 81.4% to 3.6% with DeepSeek-V3.
Under Review

Why Current AI Architectures Cannot Support Consciousness: Mediated Self-Access and the Need for Ontological Identity

Ruosen Gao and Zhou Ziheng

Under Review, INQUIRE 2026, 2026
Under Review

A Human-centric Framework for Debating the Ethics of AI Consciousness Under Uncertainty

Zhou Ziheng, Haiqiang Dai, Bin Ling, and 2 more authors

arXiv preprint arXiv:2512.02544, 2025

Abs arXiv

Proposes a three-level ethical framework for machine consciousness questions, establishing five factual determinations about AI consciousness.
ICLR

On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification

Yongliang Wu, Yizhou Zhou, Zhou Ziheng, and 7 more authors

International Conference on Learning Representations (ICLR), 2026

Abs arXiv

Proposes Dynamic Fine-Tuning (DFT) to stabilize gradient updates for SFT. A single-line code change significantly outperforms standard SFT. Accepted as ICLR 2026 Poster.
arXiv

Aligner: One Global Token is Worth Millions of Parameters When Aligning Large Language Models

Zhou Ziheng, Yingnian Wu, Song-Chun Zhu, and 1 more author

arXiv preprint arXiv:2312.05503, 2023

Abs arXiv

A parameter-efficient fine-tuning method where a single globally shared token with merely 5,000 parameters performs comparably to LoRA which requires millions of parameters.
ICCV

ARNOLD: A Benchmark for Language-Grounded Task Learning With Continuous States in Realistic 3D Scenes

Ran Gong, Jiangyong Huang, Yizhou Zhao, and 9 more authors

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

arXiv
ICLR WS

Learning to Be Fair: Modeling Fairness Dynamics by Simulating Moral-Based Multi-Agent Resource Allocation

Haiyan Feng, Yuqiao Du, Huacong Tang, and 5 more authors

ICLR 2026 Workshop AFAA, 2025

Abs

Investigates fairness as an evolving, negotiated construct using LLM-driven cognitive agents in multi-agent resource allocation simulations.