Current position
SJTU EE M.S.
M.S. student in Electrical Engineering, advised by Prof. Weinan Zhang.
Shuai Shao
M.S. student in EE at SJTU, advised by Prof. Weinan Zhang
I am a master's student at Shanghai Jiao Tong University. My research focuses on RL for LLMs and agents, agent safety and alignment, multi-agent systems, and reinforcement learning theory.
Driven by curiosity. Drawn to discovery. I do research for the joy of figuring things out: following questions that intrigue me, turning ideas into experiments, and finding something unexpected along the way.
2026 cycle
7 accepted papers
Two at ICLR, two at ICML, two at COLM (including one oral presentation), and one at NeurIPS.
Open source
1,500+ GitHub stars
Across open-source projects I have led or contributed to.
News
Recent milestones
2026 | NeurIPS
One paper accepted to NeurIPS 2026
MMSkills: Towards Multimodal Skills for General Visual Agents was accepted to NeurIPS 2026.
2026 | COLM
Two papers accepted to COLM 2026
ATBench and Rethinking Generalization in Reasoning SFT (Oral) were accepted to COLM 2026.
2026 | ICML
Two papers accepted to ICML 2026
MonoScale: Scaling Multi-Agent System with Monotonic Improvement and Are Your Agents Upward Deceivers were accepted to ICML 2026.
2026 | ICLR
Two papers accepted to ICLR 2026
Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents and Towards Self-Evolving Benchmarks: Synthesizing Agent Trajectories via Test-Time Exploration were accepted to ICLR 2026.
2025 | NeurIPS
Two papers accepted to NeurIPS 2025
RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents and AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems were accepted to NeurIPS 2025.
2025 | ICML
One paper accepted to ICML 2025
Extreme Value Policy Optimization for Safe Reinforcement Learning was accepted to ICML 2025.
Education
Academic background
Shanghai Jiao Tong University
M.S. in Electrical Engineering
Currently pursuing a master's degree in Electrical Engineering at Shanghai Jiao Tong University.
Research interests
From agent capability to agent reliability
RL for LLMs and agents, data synthesis, LLM-agent safety and alignment, and reinforcement learning theory.
Experience
Research experience
Mar 2025 - Jan 2026
Shanghai AILab
Research Intern, coadvised by Dr. Dongrui Liu and Dr. Jing Shao
- Co-first author of RiOSWorld, the first risk assessment benchmark for multimodal computer-use agents, accepted to NeurIPS 2025.
- Contributed to SafeWork-R1 on efficient reasoning RL post-training and joint safety-efficiency evaluation across large reasoning models.
- First author of the Misevolution project, a systematic study of risks in self-evolving LLM agents, accepted to ICLR 2026.
- Served as one of the student leads for AgentDoG, working on large-scale data synthesis, benchmark construction, and risk trajectory generation.
Sept 2025 - Present
Shanghai Innovation Institute
Visiting Student, advised by Prof. Weinan Zhang
- Leading MonoScale, an expansion-aware update framework for scaling LLM-based multi-agent systems under sequential agent augmentation.
- Proposed agent-conditioned task customization and natural-language memory updates to address router cold-start when adding new agents.
- Formalized sequential augmentation as a contextual bandit and derived a monotonic non-decreasing performance guarantee via trust-region memory optimization.
Oct 2024 - Sept 2025
APEX Lab, SJTU
Research Intern, advised by Prof. Weinan Zhang
- Contributed to AgentNet, a decentralized coordination framework for LLM-based multi-agent systems, accepted to NeurIPS 2025.
- Designed a retrieval-based memory system for skill refinement and agent specialization across coding, math, and QA tasks.
- Implemented self-evolving routing so agents can adapt connections and routing behavior without a central coordinator.
Mar 2024 - Oct 2024
IIOT Lab, SJTU
Research Intern, advised by Prof. Jiaxin Ding
- Co-second author on Extreme Value Policy Optimization for Safe Reinforcement Learning, accepted to ICML 2025.
- Developed an extreme value quantile optimization objective for rare, high-impact constraint violations.
- Designed extreme-priority replay and proved upper bounds on constraint violations during policy updates.
Publications
Selected Publications
ICLR 2026
ICML 2026
MonoScale: Scaling Multi-Agent System with Monotonic Improvement
Preprint 2026
Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories
Technical Report
SafeWork-R1: Coevolving Safety and Intelligence under the AI-45 Law
Technical Report
AgentDoG: A Diagnostic Guardrail Framework for AI Agent Safety and Security
COLM 2026
Oral
Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability
* indicates equal contribution. † indicates project lead.
Open Source
Selected open-source projects
Agent Safety
AgentDoG
A diagnostic guardrail framework for AI agent safety and security, built around trajectory-level risk assessment, fine-grained diagnosis, and the ATBench dataset.
Research Workflow
iDeer
A multi-source information aggregation and scheduled briefing tool that tracks research updates across platforms, ranks them with LLMs, and delivers personalized summaries by email.
Service & Additional Information
Community, talks, and background
Service
Reviewer
- ICLR 2026, ICML 2026, NeurIPS 2026, NeurIPS Responsible FM Workshop, and AAAI LaMAS Workshop.
Talks
Invited talk
- Invited talk at NICE Community
Languages
Working across Chinese, English, and French
- Chinese: Native
- English: GRE Verbal 155, Quant 170, AW 3.5
- French: DELF B2
Beyond research
Arts and athletics
- Vice principal violist in the SJTU high-level art troupe
- Main goalkeeper for the academy soccer team at SJTU
Contact
I am happy to talk about agent safety, multi-agent systems, and RL for LLMs.
The quickest way to reach me is by email. You can also find papers and updates on Google Scholar and GitHub.