🙋♀️ I am （周艺晋）, currently a second-year Ph.D. student at Shanghai Jiao Tong University and Shanghai Innovation Institute, supervised by Professor Junchi Yan and Research Scientist Jing Shao. I received my bachelor’s degree from Zhiyuan College, Shanghai Jiao Tong University, in 2024.
🔍 My research interests lie in AI agents, post-training, and trustworthiness.
💻 I am currently seeking industrial research/engineering internships related to agentic post-training. Always open to ☕️coffee chats, 💬scientific discussions, and 🤝potential collaborations—feel free to drop me an email!
Sections
Building AI Agents for Real Life.
- Agentic post-training: teaching language and vision-language models to reason, use tools, and learn from interaction.
- Trustworthy agent systems: designing safeguards and evaluations that make agents more robust against harmful propagation and unsafe behavior.
- Reliable generalization: understanding out-of-distribution behavior and building models that remain dependable beyond their training data.
🔥 News
- 2026.09: 🎉 One paper is accepted by NeurIPS 2026.
- 2026.08: 🇭🇺 Two papers are accepted by EMNLP 2026. (first author)
- 2026.06: 🇸🇪 One paper is accepted by ECCV 2026. (co-first author)
- 2026.05: 🇰🇷 One paper is accepted by ICML 2026. (first author)
- 2026.04: 🇦🇺 One paper is accepted by RSS 2026.
- 2026.03: 🎉 One paper is accepted by CVPR 2026 Findings.
- 2026.03: 🎉 One paper is accepted by CVPR 2026.
- 2025.11: 🎉 One paper is accepted by AAAI 2026.
- 2025.09: 🎓 Started my journey at Shanghai Innovation Institute.
- 2025.07: 🎓 Started a research internship at Shanghai AI Lab.
- 2025.04: 🎉 One paper is accepted by Medical Image Analysis. (first student author and corresponding author)
- 2024.06: 🎓 Graduated from SJTU and received the Excellent Bachelor’s Thesis Award (Top 1%).
- 2024.06: 🎓 Started an algorithm internship at Alibaba Cloud.
- 2024.02: 🎉 One paper is accepted by the ICML 2024 FM-Wild Workshop. (first author)
- 2023.10: 🎉 One paper is accepted by Medical Image Analysis.
🏫 Education
2025.09 - Present
Shanghai Innovation Institute
Ph.D. Candidate · Artificial Intelligence
Advisor: Professor Junchi Yan
Previously collaborated closely with Jing Shao
2024.09 - Present
Shanghai Jiao Tong University
Ph.D. Candidate · Mathematics
School of Mathematical Sciences & Institute of Natural Sciences
2020.09 - 2024.06
Shanghai Jiao Tong University · Zhiyuan College
Bachelor's degree in Mathematics and Applied Mathematics (Zhiyuan Honors Program)
Ministry of Education's Pilot Program for Training Top Students in Basic Disciplines
Zhiyuan College admits students from the university-wide top 10%.
2017.09 - 2020.06
Hangzhou Xuejun High School
Hangzhou, China
Always happy to connect with fellow high school alumni🤝
💼 Research Experience
2025.07 - 2026.07
Shanghai Artificial Intelligence Laboratory
Research Intern
Jing Shao's team
Research focus: AI agents & trustworthy AI
2024.06 - 2024.08
Alibaba Cloud Computing Co., Ltd.
Algorithm Intern
Transformer-based time series analysis
📃 Publications
Research Trajectory
Selected first- and co-first-author works.
arXiv v1
Accepted
Post-training
Trustworthy
EMNLP 2026
Teaser
Tool-calling decision failures in LLM-based agents.
Method
TRUST combines uncertainty-aligned reward with turn-level decision calibration.
Exploring Agentic Tool-Calling Decisions via Uncertainty-Aligned Reinforcement Learning
TRUST
Empirical Methods in Natural Language Processing (EMNLP), 2026
arXiv v1: 2026.06 · Accepted: 2026.08
Claim: TRUST addresses suboptimal tool-use decisions in LLM agents, including unsupported tool calls and hallucinated direct responses. It adds uncertainty quantification as a repulsive force in reward design and uses lightweight key-turn annotations for multi-turn post-training; experiments report improved decision and agent performance with more reliable uncertainty estimates.
EMNLP 2026
Teaser
Existing multi-agent safeguards compared with infection-aware safeguarding.
Method
Overview of the proposed INFA-GUARD method.
INFA-GUARD: Mitigating Malicious Propagation via Infection-Aware Safeguarding in LLM-Based Multi-Agent Systems
INFA-Guard
Empirical Methods in Natural Language Processing (EMNLP), 2026
arXiv v1: 2026.01 · Accepted: 2026.08
Claim: INFA-GUARD models benign, attack, and infected agents as distinct categories in multi-agent systems. Infection-aware detection and topological constraints localize attack sources and infected ranges; remediation replaces attackers and rehabilitates infected agents, with the reported experiments reducing attack success rate by an average of 33%.
ECCV 2026
Teaser
Context-guided chain-of-thought for identifying implicit contextual risks.
Method
Two-stage HomeGuard training pipeline and sequential risk reasoning.
HomeGuard: VLM-based Embodied Safeguard for Identifying Contextual Risk in Household Task
HomeGuard
European Conference on Computer Vision (ECCV), 2026
arXiv v1: 2026.03 · Accepted: 2026.07
Claim: HomeGuard is an architecture-agnostic safeguard for contextual risks in household tasks. Its Context-Guided Chain-of-Thought performs active perception of interaction targets and spatial neighborhoods before semantic judgment, supported by a grounding dataset and process-reward reinforcement fine-tuning; the paper reports over 30% higher risk-match rates than base models with reduced oversafety.
ICML 2026
Teaser
Outlier diversity improves OOD detection.
Method
GROD generates synthetic OOD data during fine-tuning and guides learning with an ID-OOD classifier.
How Out-of-Distribution Detection Learning Theory Enhances Transformer: Learnability and Reliability
OOD
International Conference on Machine Learning (ICML), 2026
arXiv v1: 2025.02 · Accepted: 2026.05
Claim: This paper develops PAC learning theory for OOD detection with Transformers, characterizing conditions under which outliers are learnable and distinguishable. It derives a soft-synthetic-outlier training strategy from the theory and reports state-of-the-art performance across data formats.
ICML 2024 Workshop
Teaser
Overview of the GROD algorithm.
Method
Computational-cost comparison among OOD detection methods.
GROD: Enhancing Generalization of Transformer with Out-of-Distribution Detection
GROD
ICML 2024 FM-Wild Workshop
arXiv v1: 2024.06 · Accepted: 2024.06
Claim: GROD improves Transformer generalization to out-of-distribution data by generating synthetic outliers and penalizing their misclassification. The paper establishes an OOD-detection PAC learning theory for Transformers and reports state-of-the-art results across NLP and computer-vision OOD tasks.
ArXiv preprint
Teaser
Physical-link defense versus memory-link defense.
Method
Memory-link poisoning and MAPLE-Guard defense overview.
MAPLE-Guard: Memory-Aware Link Enforcement Against Memory-Link Poisoning in Multi-Agent Systems
MAPLE-Guard
ArXiv preprint
arXiv v1: 2026.08 · Under Review
Claim: MAPLE-Guard protects memory-enabled multi-agent systems against poisoning that propagates through private and shared memories. It gates memory writes, retrieval, promotion, and cross-agent reuse to quarantine risky memories and block poisoned private memories from entering shared memory; the paper reports large ASR reductions on LongMemEval and AppWorld.
Medical Image Analysis
Teaser
Learnable stain normalization with multiple color-space fusion.
Method
Self-attention feature aggregation and transformation across color spaces.
Learnable color space conversion and fusion for stain normalization in pathology images
LStainNorm
Medical Image Analysis (IF 10.9)
Published: 2025.04
Claim: LStainNorm introduces a learnable stain-normalization layer that learns optimal stain characteristics without manual template selection and adds self-attention to fuse multiple color spaces. Across classification and nuclei-segmentation datasets, the paper reports gains over conventional and GAN-based methods, faster end-to-end inference, and an interpretable learned template.
Co-authored Papers9
ArXiv preprint
Teaser
Navi-Agent overview.
Method
Navi-Agent system overview with typed sub-goal decomposition and visual memory.
Navi-Agent: Unlocalized Monocular Navigation Agent
Navi-agent
ArXiv preprint
arXiv v1: 2026.09 · Under Review
Claim: Navi-Agent builds a coordinate-free visual spatial state for zero-shot VLN-CE from observations and executed motion histories. Its visual-anchor graph supports approximate self-localization, progress verification, and visual backtracking recovery, with experiments reporting state-of-the-art performance among geometry-constrained methods.
ArXiv preprint
Teaser
Overview of the MVP-Nav navigation system.
Method
Overview of the MVP-Nav framework and its recursive perception-planning loop.
MVP-Nav: Multi-layer Value Map Planner Navigator
MVP-Nav
RSS 2026
arXiv v1: 2026.06 · Accepted: 2026.06
Claim: MVP-Nav is an RGB-only object-goal navigation framework that reconstructs physical occupancy from monocular observations using 3D foundation models. Its Multi-layer Value Map combines semantic priorities with reconstructed geometry for physically grounded planning, and experiments report state-of-the-art performance among depth-free methods.
CVPR 2026
Teaser
Geometrically-constrained agent for spatial reasoning.
Method
Overall GCA paradigm with task formalization and constrained geometric computation.
Geometrically-Constrained Agent for Spatial Reasoning
GCA
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
arXiv v1: 2025.11 · Accepted: 2026.02
Claim: GCA addresses the semantic-to-geometric gap in VLM spatial reasoning with a training-free two-stage agent. The VLM first translates an ambiguous query into a formal verifiable constraint, then executes tool calls within those deterministic bounds; experiments report roughly 27% gains over training-based and tool-integrated baselines.
CVPR 2026 Findings
Teaser
JANO predicts regional convergence from early-stage complexity and allocates computation adaptively.
Method
Adaptive token scheduling through an interleaved generation pipeline.
JANO: Adaptive Diffusion Generation with Early-stage Convergence Awareness
Jano
CVPR 2026 Findings
arXiv v1: 2026.02 · Accepted: 2026.02
Claim: JANO accelerates Diffusion Transformer generation by recognizing regional convergence requirements in early denoising steps and adaptively scheduling tokens. The training-free method reports an average 2.0× speedup, up to 2.4×, while preserving generation quality.
Medical Image Analysis
Teaser
Overall architecture of ClusterSeg with a CNN-transformer hybrid encoder and composite decoder.
Method
Partially supervised PS-ClusterSeg training framework.
ClusterSeg: A crowd cluster pinpointed nucleus segmentation framework with cross-modality datasets
ClusterSeg
Medical Image Analysis (IF 10.9)
Published: 2023.10
Claim: ClusterSeg targets difficult instance segmentation of crowded nuclei with a convolutional-transformer encoder and a 2.5-path decoder for masks, contours, and clustered edges. The paper also introduces an annotation-efficient clustered-edge strategy and PS-ClusterSeg, evaluated across multiple microscope, cytopathology, and histopathology modalities.
Open-source toolkit
Teaser
DeepSafe configuration-driven architecture.
Method
DeepScan configuration-driven architecture.
DeepSight: An All-in-One LM Safety Toolkit
DeepSight
Open-source safety toolkit
arXiv v1: 2026.02
Claim: DeepSight unifies large-model safety evaluation and diagnosis through two open-source toolkits, DeepSafe and DeepScan. By sharing task and data protocols, it connects black-box behavioral evaluation with white-box internal diagnosis and supports frontier-AI risk evaluation.
Technical report
Teaser
Report abstract: five critical dimensions of frontier AI risk.
Method
Report contents: model information, frontier risk evaluations, and conclusions.
Frontier AI Risk Management Framework in Practice: A Risk Analysis Technical Report
FrontierAIRisk
Frontier AI Risk Management Framework in Practice, Version 1.5
arXiv v1: 2026.02
Claim: This technical report updates a frontier-AI risk analysis framework across cyber offense, persuasion and manipulation, strategic deception, uncontrolled AI R&D, and self-replication. It adds new high-fidelity scenarios and mitigation strategies, including risks from agent memory/tool expansion and resource-constrained self-replication.
NeurIPS 2026
Teaser
CUBIC achieves Pareto-optimal quality-efficiency trade-offs across three generation tasks.
Method
CUBIC converts curvature estimates into adaptive partitions, cache intervals, and sparse execution.
CuBic: Curvature-Driven Dynamic Inference Caching for Fast, High-Fidelity Flow Matching
CuBic
NeurIPS 2026
Accepted: 2026.09
Claim: CUBIC uses trajectory curvature to adapt spatial partitioning and cache update intervals for Flow Matching. Its lightweight endpoint-drift proxy enables train-free dynamic caching; experiments across image, image-editing, and video generation report Pareto improvements with an average 2.5x speedup.
Paper not public yet Code not public yet
AAAI 2026
Teaser
Interactive versus termination-oriented safety evaluation for embodied agents.
Method
Evaluation scenario generation in IS-Bench.
IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks
IS-Bench
AAAI Conference on Artificial Intelligence (AAAI), 2026
arXiv v3: 2025.06 · Accepted: 2025.11
Claim: IS-Bench evaluates interactive safety rather than only terminal task outcomes, using 161 simulated household scenarios containing 388 safety risks. Its process-oriented evaluation checks whether mitigation actions occur in the correct order; experiments find current VLM agents lack interactive safety awareness and that safety-aware chain-of-thought can trade off task completion.
🏆 Awards
Shanghai Jiao Tong University
- Excellent Bachelor's Thesis (Top 1%)2024.06
- “Chun-Tsung Scholar” Honorary Title, conferred by Nobel Laureate Tsung-Dao Lee2023.05
- Outstanding Graduate2024.06
- Merit Scholarship2021.11
- Zhiyuan Honor Scholarship2020 · 2021 · 2022
- Undergraduate Excellence Scholarship2021 · 2022 · 2023
🏸 Interests
Badminton
Aside from research, I am a passionate badminton player with over a decade of experience. Before entering university, I received systematic training and represented my prefecture-level city in a series of competitions, including the 15th Zhejiang Provincial Games, winning titles such as the urban district singles and doubles champion, 5th place in singles at the provincial level, and 3rd place in doubles. During my university years, I was a member of the Shanghai Jiao Tong University varsity team, serving as a key starting player and securing honors including 7th place in singles in Shanghai, team champion, and champion of the SJTU General Sports Association Cup. Currently, I primarily play for the joy of the sport while continuing to participate in amateur tournaments, including winning the women's singles championship in the tournament organized by Shanghai AI Lab during my internship, and placing 5th in women's doubles at the YONEX "Meiwuzhi" tournament.
✈️ Academic Conference Attendance
Academic conferences are one of the best parts of my Ph.D. journey!
📌 ICML 2026 · Seoul
📌 RSS 2026 · Sydney
📌 ECCV 2026 · Malmö
🇭🇺 See you in EMNLP 2026, Budapest!
🔗 Academic Friends
My academic friends include Xiaoya Lu, Wenyuan Xie, Linqian Zeng, Yuyang Chen, and Qi Chen.
Many more friends have shaped my journey—if I missed you here, please send me a friendly reminder 🌹