Beyond Scaling Laws: Yaqing Wang on Data-Efficient Learning as AI’s Next Frontier (IJCAI 2026)

Scaling laws have become a de‑facto consensus in the large‑model era, but the community is confronting multiple bottlenecks: high‑quality annotation data are running out, compute costs are exploding, and black‑box models remain opaque.

Many real‑world tasks naturally lack massive datasets—drug discovery, cold‑start recommendation, and scientific exploration are cited as examples where data scarcity is a fundamental property rather than a temporary technical shortfall.

At IJCAI 2026’s Early Career Spotlight, BIMSA associate professor Yaqing Wang (PhD, HKUST; former senior researcher at Baidu) presented a systematic response. Her talk, titled “Toward Data‑Efficient Intelligence: From Few‑Shot Learning to Agentic Systems,” traced the evolution from few‑shot learning, meta‑learning, and contextual learning to the newly proposed Data‑Efficient Agentic Learning (DEAL) framework.

Core Insight: Structured Priors Enable Data‑Efficient Generalization

Wang argues that human intelligence generalizes from extremely few examples thanks to three layers of prior knowledge—genetic encoding, social‑cultural background, and individual experience. She demonstrates theoretically that, under the studied tasks and model settings, In‑Context Learning (ICL) can be interpreted as a data‑dependent form of meta‑learning, with the Transformer’s self‑attention implicitly internalizing the adaptation process.

Empirical results from ICLR 2025 show that, on linear regression and classification benchmarks, ICL matches or exceeds mainstream meta‑learning methods (optimization‑based, metric‑based, and amortization‑based) when parameters are appropriately chosen, and even dynamically selects the best adaptation rule for mixed task families.

Three‑Level Prior Integration

Data level: Use prior rules to generate or augment samples, filling sparse data spaces.

Model level: Introduce inductive bias (e.g., convolution for images) to constrain hypothesis‑space complexity.

Algorithm level: Leverage priors to guide search in parameter space, improving initialization and gradient direction.

In the agentic setting, data D = D_loop ∪ D_ext combines external annotations with the agent’s own interaction trajectories (environment feedback, actions, rewards, failures). These loops are costly to collect and subject to distribution shift, creating a “data‑silo” bottleneck.

DEAL Framework

To build strong agents under strict data budgets, Wang and collaborators propose three technical directions:

Experience Augmentation: Generate and transform experiences to increase information density and diversity in limited interaction traces.

Agent Structural Design: Create modular, autonomous workflows and knowledge‑reuse mechanisms to reduce reliance on human supervision.

Learning Paradigm: Flexibly choose between targeted ICL optimization, few‑shot fine‑tuning, or budget‑efficient reinforcement learning based on the available budget.

This framework eliminates the need for massive human trajectory datasets, enabling high‑performing agents with limited data.

Structured Priors for AI for Science and Personalized Agents

Two focal applications are highlighted:

AI for Science: Incorporate rigorous mathematical formulas and physical laws as priors in domains such as chemistry, physics, and biomedicine, improving interpretability and stability while compensating for scarce wet‑lab data.

Industrial‑level personalized agents: Inject human behavior priors into education or embodied‑robot scenarios, allowing rapid cold‑start adaptation to individual user preferences.

Wang also helped launch the IEEE CIS NNTC Task Force on Data‑Efficient Agentic Learning (TF‑DEAL), chaired by Yao Quan‑Ming (Tsinghua) with an international advisory board, to foster benchmarks, tutorials, and community collaboration.

References:

Y. Wang, Q. Yao, J. T. Kwok, L. M. Ni. “Generalizing from a Few Examples: A Survey on Few‑Shot Learning.” ACM Computing Surveys, 2020.

Y. Wang et al. “Property‑Aware Relation Networks for Few‑Shot Molecular Property Prediction.” NeurIPS, 2021.

Y. Wang et al. “Warming Up Cold‑Start CTR Prediction by Learning Item‑Specific Feature Interactions.” KDD, 2024.

S. Wu, Y. Wang, Q. Yao. “Learning to Learn with Contrastive Meta‑Objective.” NeurIPS, 2025.

H. Nie, Y. Wang et al. “Adaptive Preference Arithmetic: Modeling Dynamic Preference Strengths for LLM Agent Personalization.” NeurIPS, 2025.

S. Wu, Y. Wang, Q. Yao. “Why In‑Context Learning Models are Good Few‑Shot Learners?” ICLR, 2025.

S. Wu, Y. Wang, Q. Yao. “Searching to Modulate for Cold‑Start Recommendation.” TPAMI, 2026.

Y. Tan, Q. Yao, Y. Wang. “DGNet: Discrete Green Networks for Data‑Efficient Learning of Spatiotemporal PDEs.” ICLR, 2026.