Hongyi Du
MSCS student at UIUC
U LabProf. Jiaxuan You’s research group
Where does
intelligence
live over time?
I study how intelligence moves beyond models and individual agents into relationships, organizations, and institutions—structures that can preserve, revise, and transmit collective capability across time.
From agents, to organizations, to longer-lived forms of collective intelligence.
Hongyi Du杜洪一
I’m an MSCS student at the University of Illinois Urbana-Champaign. I’m fortunate to be a member of U Lab and to work with Prof. Jiaxuan You on AI agents, their organizations, and the interfaces that connect them to people.
Intelligence is a beginning.
What makes it persist?
A model can contain capability. An agent can turn it into action. But contexts end, members are replaced, and organizations change or disappear. I ask how intelligence becomes embedded in structures that outlast the individuals that created it.
Intelligence can migrate into structures whose lifetime exceeds that of its current carriers.
THE RESEARCH NOTE BEHIND THIS AGENDAWhere Does Intelligence Live Over Time?From agent memory to organizational learning and civilizational inheritance.A research space for
persistent intelligence.
How intelligence persists.
Where can intelligence live, and what must it outlast? Not every combination defines a distinct research object.
From individual capability to collective history.
Nine main stops connect the research questions. This is a reading route, not a ladder of completed achievements.
Explore the research questions.
Select a name to travel in the scene above. Open its project or note to read more.
| Y ↑ / X → | 0Individual | 1Group | 2Organization | 3Society | 4Civilization |
|---|---|---|---|---|---|
| 4Across Successive Generations | |||||
| 3Across System Reconfiguration | |||||
| 2Across Carrier Turnover | |||||
| 1Across Episodes | |||||
| 0Within Interaction |
Coordinates organize questions, not measured progress. Proposed directions are provisional, not completed results.
Recent papers from my research.
Agentic Sociology develops the framework for organizations and societies. RELIC studies persistent organizational capability. RINI examines contribution attribution against prior work.
Selected work.
RELIC
Governed protocol formation, runtime binding, and capability that persists across work and member turnover.
RINI
Auditing whether generated research proposals correctly attribute established contributions, then repairing unsupported novelty claims without erasing the research question.
ProtocolBench
Comparing A2A, ACP, ANP, and Agora across success, latency, overhead, and failures—with learned selection through ProtocolRouter.
MultiAgentBench
MARBLE’s interactive environments evaluate task outcomes and coordination milestones across different communication structures and planning strategies.
Questions I’m
exploring.
Where should collective experience be stored?
02How does an organization know what to preserve, revise, or forget?
03How can organizational capability survive member and structural change?
↗What survives when the organization itself does not?
FRONTIERWhat persists when
the members change?
Governed executable protocols provide a concrete starting point for a wider research program on organizational learning.
Follow the researchHow intelligence
learns to persist.
From Agent Organization to Civilizational Intelligence
Organizational learning, evolution, death, succession, and lineage: what happens when artificial intelligence has not only memory, but history?
Where Does Intelligence Live Over Time?
From agent memory to organizational learning and civilizational inheritance: a research program for intelligence that outlives its original carriers.
Why agent cooperation fails.
When more agents create more coordination work—not better work. From MultiAgentBench to persistent organizational capability.
AI agents are still in the organizational stone age.
Modern individual capabilities, primitive working arrangements: the missing language, institutions, and continuity of collective work.
Protocols are a foundation for agent society.
From communication protocols to generated, governed rules of work—and the institutions between organizations.
Three levels of human–organization interaction.
Direct participation, a transparent liaison, and liaison-first interaction: the P1–P3 design progression in the RELIC appendix.
Research papers.
RELIC: From Multi-Agent Collaboration to Persistent Organizational Capability
BibTeX
@misc{relic,
title = {RELIC: From Multi-Agent Collaboration to Persistent Organizational Capability},
author = {Hongyi Du and Tianyi Zhang and Weijia Zhang and Yi Yang and Haofei Yu and Kunlun Zhu and Tianxiang Dai and Shang Jiang and Zhelun Gao and Jiaxin Pei and Shang Zhu and Jiaxuan You},
year = {2026},
note = {Preprint},
eprint = {2609.32965},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.32965}
}RINI: Seeing the Prior Is Not Enough
BibTeX
@misc{rini,
title = {RINI: Seeing the Prior Is Not Enough},
author = {Hongyi Du and Tianyi Zhang and Heng Wang and Zhelun Gao and Yimei Liu and Ambrose Luo and Annie Hao and Jiayan Ni and Jiawei Han and Jiaxuan You},
year = {2026},
note = {Preprint},
eprint = {2609.33284},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.33284}
}ProtocolBench: Which LLM MultiAgent Protocol to Choose?
BibTeX
@misc{protocolbench,
title = {ProtocolBench: Which LLM MultiAgent Protocol to Choose?},
author = {Hongyi Du and Jiaqi Su and Jisen Li and Lijie Ding and Yingxuan Yang and Peixuan Han and Xiangru Tang and Kunlun Zhu and Jiaxuan You},
year = {2026},
note = {ICML 2026},
eprint = {2510.17149},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2510.17149}
}Ψ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues
BibTeX
@misc{psibench,
title = {Ψ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues},
author = {Peixuan Han and Hongyi Du and Jiayu Liu and Yihang Sun and Yutong Liu and Jiaxuan You},
year = {2026},
note = {Preprint},
eprint = {2606.02754},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2606.02754}
}MultiAgentBench: Evaluating the Collaboration and Competition of LLM Agents
BibTeX
@misc{multiagentbench,
title = {MultiAgentBench: Evaluating the Collaboration and Competition of LLM Agents},
author = {Kunlun Zhu and Hongyi Du and Zhaochen Hong and Xiaocheng Yang and Shuyi Guo and Zhe Wang and Zhenhailong Wang and Cheng Qian and Xiangru Tang and Heng Ji and Jiaxuan You},
year = {2025},
note = {ACL 2025},
eprint = {2503.01935},
archivePrefix = {arXiv},
url = {https://aclanthology.org/2025.acl-long.421/}
}ModelingAgent: Bridging LLMs and Mathematical Modeling for Real-World Challenges
BibTeX
@misc{modelingagent,
title = {ModelingAgent: Bridging LLMs and Mathematical Modeling for Real-World Challenges},
author = {Cheng Qian and Hongyi Du and Hongru Wang and Xiusi Chen and Yuji Zhang and Avirup Sil and Chengxiang Zhai and Kathleen McKeown and Heng Ji},
year = {2025},
note = {Findings of EMNLP 2025},
eprint = {2505.15068},
archivePrefix = {arXiv},
url = {https://aclanthology.org/2025.findings-emnlp.85/}
}EscapeBench: Towards Advancing Creative Intelligence of Language Model Agents
BibTeX
@misc{escapebench,
title = {EscapeBench: Towards Advancing Creative Intelligence of Language Model Agents},
author = {Cheng Qian and Peixuan Han and Qinyu Luo and Bingxiang He and Xiusi Chen and Yuji Zhang and Hongyi Du and Jiarui Yao and Xiaocheng Yang and Denghui Zhang and Yunzhu Li and Heng Ji},
year = {2025},
note = {ACL 2025},
eprint = {2412.13549},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2412.13549}
}From my research.
RELIC: building long-horizon agent organizations
A long-horizon testbed for LLM-agent organizations, including OrgEnv and professional profile graphs.
Read updateJoining the UIUC ATLAS Moonshot
Working on graph-grounded learning and reskilling, learner state, and human verification.
Read updateProtocolBench accepted to ICML 2026
Our work studies multi-agent communication protocols and scenario-aware selection with ProtocolRouter. First-author paper.
Read update
Loading comments…