ACL 2026
Chain-of-Memory: Lightweight Memory Construction with Dynamic Evolution for LLM Agents
A lightweight memory construction mechanism with dynamic evolution, giving LLM agents persistent and continually refined memory at low overhead.
About
I am an Associate Professor at the Research Center for Cognitive Intelligence Systems and the State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS), where I also serve as a master’s supervisor. From August 2025 to August 2026 I was a visiting scholar at the NExT++ Research Centre, National University of Singapore, hosted by Prof. Tat-Seng Chua. I received my Ph.D. from ICT, CAS in 2021 under the supervision of Prof. Xueqi Cheng and Prof. Huawei Shen, and received my B.S. from Harbin Institute of Technology (Weihai) in 2016. My earlier work centers on graph representation learning, in particular spectral graph neural networks, with applications in financial risk control and social computing. My current research focuses on large language models and LLM agents — inference-time alignment, process reward modeling, memory and planning for agents, and multi-agent social simulation.
Openings
We are looking for students (third/fourth year undergraduates or master's students) as interns to collaborate with us in research, as well as assistant professors and postdocs working on large language models and LLM agents. If you are interested, please feel free to contact me at any time.
Selected Publications
All publications →
ACL 2026
ACL 2026
Learning from Mistakes: Negative Reasoning Samples Enhance Out-of-Domain Generalization
Negative reasoning samples — the wrong chains models usually discard — turn out to carry the signal that drives out-of-domain generalization.
ACL 2026
HAG: Hierarchical Demographic Tree-based Agent Generation for Topic-Adaptive Simulation
A hierarchical demographic tree generates agent populations that adapt to the topic under discussion, making social simulation both faithful and controllable.
ACL 2026
ROMA: Real-time Omni-Multimodal Assistant with Interactive Streaming Understanding
An omni-multimodal assistant that understands audio-visual streams as they arrive, so it can respond mid-stream instead of waiting for the input to finish.
ACL 2026
Do We Always Need Query-Level Workflows? Rethinking Agentic Workflow Generation for Multi-Agent Systems
Query-level workflow generation costs far more than task-level yet barely helps — a single well-optimized workflow already covers most queries.
EACL 2026
Incentivizing Strong Reasoning from Weak Supervision
We show that a weak teacher — far smaller and less performant than the student — can substantially boost student reasoning, without expert supervision or costly RL.
WSDM 2026
Multi-Personality Generation of LLMs at Decoding-Time
Multiple personality traits are composed at decoding time, letting one LLM express a target persona without any personality-specific training.
NeurIPS 2025
Inference-Time Alignment in Continuous Space
We reformulate alignment as an iterative optimization over an energy function on logits in continuous space, defined by the optimal RLHF policy — achieving deep alignment without retraining.
ACM MM 2025
MIGE: Mutually Enhanced Multimodal Instruction-Based Image Generation and Editing
A unified framework that treats instruction-based image generation and editing as one task, letting the two settings reinforce each other instead of being trained apart.
Recent News
All news →- 2026-08 Completed my one-year visiting scholarship at the NExT++ Centre, National University of Singapore.
- 2026-05 Five papers accepted at ACL 2026 — three in the Main Conference (Chain-of-Memory, Learning from Mistakes, HAG) and two in Findings (ROMA, Rethinking Agentic Workflow Generation).
- 2025-12 One paper accepted at WSDM 2026: Multi-Personality Generation of LLMs at Decoding-Time.
- 2025-09 One paper accepted at NeurIPS 2025: Inference-Time Alignment in Continuous Space.
- 2025-08 Started a one-year visit at the NExT++ Centre, National University of Singapore, hosted by Prof. Tat-Seng Chua.