How can AI systems scale when direct supervision is limited? This half-day tutorial examines settings with scarce labels, weak or delayed feedback, and unstable self-generated supervision. It presents a technically grounded framework spanning data-efficient priors, graph & relational structure as implicit supervision, and adaptive agentic systems — with emphasis on methods, evaluation, and open challenges broadly relevant to the IJCAI community.
🎓 Target Audience
Researchers & practitioners in machine learning, foundation models, graph learning, and multi-agent systems; PhD students and postdocs seeking a structured entry into limited-supervision methods; expert non-specialists from NLP, data mining, AI for science, and web/social computing. Graduate-level familiarity with ML and neural networks is expected; no prior background in graph learning, instruction tuning, or agentic systems is assumed.
✨ Distinctive Contribution
A single narrative arc from priors → structure → interaction, enabling attendees to compare methods often studied separately and understand shared challenges in stability, evaluation, error propagation, and trustworthy deployment. By the end, participants gain a clear conceptual map, practical guidance on when to apply different limited-supervision strategies, and a research agenda of open questions.
📌 Two-Sentence Description (Brochure Version):
This half-day tutorial examines how AI systems can scale when direct supervision is limited, including settings with scarce labels, weak or delayed feedback, and unstable self-generated supervision. It presents a unified and technically grounded framework spanning data-efficient priors, graph and relational structure as implicit supervision, and adaptive agentic systems, with emphasis on methods, evaluation, and open challenges broadly relevant to the IJCAI community.
👥 Presenters
🎙️
Song Zixing
Lecturer (UK Assistant Professor), University of Bristol
Zixing Song is currently a Lecturer (UK Assistant Professor) at the University of Bristol. Prior to that, he was a Postdoctoral Research Associate at the University of Cambridge. He received his Ph.D. in Computer Science and Engineering from The Chinese University of Hong Kong. He also obtained his B.Eng. in Computer Science from Southeast University in China. His research centers on data-centric graph machine learning and its application to drug design and broader AI4Science domains, driven by a vision of developing data-efficient, agentic, and trustworthy models for social good. Currently, he serves as Vice Chair of the IEEE Computational Intelligence Society (CIS) Task Force on Data-efficient Agentic Learning. He has also received several international awards, including the AAAI New Faculty Highlight, the Junior Research Fellowship at Wolfson College Cambridge, and the INNS Doctoral Dissertation Award.
🎙️
Wang Yaqing
Associate Researcher, Beijing Institute of Mathematical Sciences and Applications (BIMSA)
Yaqing Wang is an Associate Professor at the Beijing Institute of Mathematical Sciences and Applications (BIMSA). She received her Ph.D. in Computer Science and Engineering from the Hong Kong University of Science and Technology, advised by Prof. Lionel M. Ni and Prof. James T. Kwok. Her research focuses on machine learning and artificial intelligence, with an emphasis on data-efficient generalization, including few-shot learning, in-context learning, and adaptive agents. Dr. Wang has published 37 papers in leading venues such as NeurIPS, ICML, ICLR, KDD, TheWebConf, TPAMI, JMLR, and TIP, with 6000 citations. She serves as an Associate Editor of Neural Networks, an editorial board member of Machine Learning, and an Area Chair for ACL Rolling Review. Her techniques have been deployed in large-scale real-world systems at Baidu, Meituan, and other industry applications. She is a recipient of the AAAI New Faculty Highlight Program and the Beijing Nova Program, and is listed among the World's Top 2% Scientists in 2024 and 2025.
🎙️
Yao Quanming
Associate Professor, Tsinghua University
Dr. Quanming Yao, Associate Professor, Department of Electronic Engineering, Tsinghua University. Previously at 4Paradigm Inc., promoted from Researcher to Senior Scientist, founder and leader of its machine learning research team. Ph.D., HKUST; B.Eng., HUST.
IEEE Top 30 Early Career Researchers to Watch; inaugural Intech Prize, Ant Group; Aharon Katzir Young Investigator Award, INNS; Forbes 30 Under 30 China; Google Fellowship. Invited early-career talks at AAAI and IJCAI.
Area Chair for ICML, NeurIPS, and ICLR; Tutorial Chair, IJCAI 2025; Program Co-chair, ADMA 2026; Action Editor, IEEE TPAMI; Senior Action Editor, Neural Networks. Fellow of IET; Young Fellow of BAAI.
📌 Together, the presenters offer an integrated view across priors, structure, and interaction.
⏱️ Tutorial Schedule (Half-day, includes 25-minute break)
Experience graphs for data collection; decision graphs for action selection and test-time search; learning graphs for feedback propagation and credit assignment
Zixing Song
10:45 - 11:10
☕ Break & Networking
11:10 - 11:55
🤖 Module C: Adaptive Agents under Limited/Delayed Feedback
External vs. internal feedback bottlenecks, test-time structural adaptation, multi-agent coordination, disagreement/critique/routing as surrogate supervision, stable self-evolution and safeguards against collapse/overconfidence/drift
Quanming Yao
11:55 - 12:30
🧠 Synthesis, Discussion & Q&A
Decision-oriented comparison, evaluation principles, trustworthy scaling, open research challenges, and audience discussion
All (moderated by Yao)
※ Total duration: 3 hours 30 minutes including a 25-minute break. Each module blends core principles, representative methods, practical design trade-offs, and open questions.
Graph structure improves agentic learning by increasing the information extracted from scarce labels, rollouts, or interactions. This part of the tutorial organizes methods into three operational lenses. Experience graphs shape data collection by constraining synthesis and exposing intermediate supervision. Decision graphs guide action selection by enabling reusable plans and limited test-time search. Learning graphs propagate feedback by aggregating rollouts and assigning credit from sparse outcomes. We clarify each lens in terms of nodes, edges, and where graph operations act, and discuss representative works.
⚖️ Responsible AI & Key Risks
The tutorial explicitly addresses risks arising under limited supervision: bias amplification from scarce labels, error propagation in pseudo-labeling/self-training, spurious correlations in graph data, and instability in self-improving agent loops. We emphasize practical safeguards: confidence calibration, selective human oversight, robust evaluation under weak supervision, and transparent reporting of failure modes and domain limits. These issues are especially critical in scientific, biomedical, and socially consequential settings — limited supervision must never be confused with reduced responsibility. Trustworthy scaling is framed as both a technical and scientific question: improving capability while preserving reliability, transparency, and appropriate human control.
📎 Reference List
Generalizing from a Few Examples: A Survey on Few-Shot Learning. ACM Computing Surveys, 2020.
Finetuned Language Models Are Zero-Shot Learners. ICLR, 2022.
Why In-Context Learning Models are Good Few-Shot Learners? ICLR, 2025.
No Change, No Gain: Empowering Graph Neural Networks with Expected Model Change Maximization for Active Learning. NeurIPS, 2023.
Beyond Scaling: A Survey on Data-Efficient Agentic Learning. IJCAI, 2026.
Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS, 2023.
Language Model Networks: Supervision-Efficient Learning through Dense Communication. ICML, 2026.
For inquiries regarding the tutorial, please contact the organizers: Quanming Yao (Tsinghua University) qyaoaa@tsinghua.edu.cn. Further details and live updates will be available on the official IJCAI-ECAI 2026 conference website.