2026 3rd International Conference on Artificial Intelligence and Natural Language Processing (AINLP 2026)

Hai Zhao is a Tenured Professor and Ph.D. Supervisor at the School of Computer Science and Engineering, Shanghai Jiao Tong University (SJTU), where he also serves as Director of the Institute of Artificial General Intelligence (IAGI). His research focuses on natural language processing (NLP) and foundational deep learning methodologies.

He has authored over 200 academic publications, including more than 100 papers in CCF A/B-tier conferences and 20+ papers in Chinese Academy of Sciences (CAS) Q1 SCI-indexed journals, among which four appear in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI). His work has garnered over 15,000 Google Scholar citations. He currently serves as a Member of the Technical Committee on Chinese Information Processing under the China Computer Federation (CCF), and Deputy Director of the Artificial Intelligence Technical Committee of the Shanghai Computer Society.

In service to the academic community, he held the role of Senior Area Chair for the Linguistic Analysis, Morphology, and Word Segmentation tracks at ACL 2017–2019, acts as Executive Editor for both the ACL Rolling Review (ARR) and Transactions of the Association for Computational Linguistics (TACL), and has served as (Senior) Program Committee Member for AAAI, IJCAI, and NeurIPS in recent years. He has led top-performing systems on major international NLP benchmarks including RACE, SQuAD 2.0, HotpotQA, and HellaSwag, with his team being the first to develop an NLU system that surpasses human performance on several of these tasks.

In 2023, he published Natural Language Understandingwith Tsinghua University Press—the first Chinese textbook-monograph hybrid focused on large language models. He leads the development of the BatGPT series of LLMs, which have been deployed in vertical industrial applications and adopted by multiple national agencies; notably, he built Shanghai’s first government-procured LLM application system. He also pioneered BriLLM, the first brain-inspired large language model to depart from conventional machine learning paradigms. Named an Elsevier Highly Cited Researcher for four consecutive years (2022–2025).

Title: Scaling Law, A Fluke, Not a Path: A Mathematical Theory of AGI and its Brain-inspired LLM Exploration

Abstract: Current AGI research is fragmented across capability-, mechanism-, and endogeneity-focused paradigms, with most industrial roadmaps over-relying on unsupervised scaling of large language models (LLMs)—a path lacking rigorous mathematical grounding. We demonstrate that pure symbolic LLM systems face three insurmountable barriers: a topological gap preventing discrete symbols from capturing high-dimensional continuous physical manifolds, a causal gap blocking progression beyond Pearl’s associational tier using observational data alone, and a thermodynamic gap rendering perfect prediction physically impossible under Bekenstein and Bremermann limits. Scaling also faces structural economic collapse: marginal performance gains decay asymptotically as training costs grow superlinearly.

To resolve these limitations, we derive necessary and sufficient conditions for computable AGI: an endogenous compression-driven self-consistency loop enabling autonomous convergence without external input, and exogenous physical grounding via perception-action coupling that aligns internal fixed points to real-world states. We further prove the mandatory existence of a MetaPred self-referential prediction module supporting infinite recursive metacognition, with direct empirical alignments to cortical semantic mapping and neural oscillation mechanisms in biological brains.

For practical deployment, we propose the Signal-Fully-connected Flow (SiFu) non-representational learning paradigm and the BriLLM-MetaPred architecture, which delivers brain-inspired AGI capabilities on commodity GPU hardware without relying on spike neural network infrastructure. This work provides the first mathematically falsifiable framework for AGI development, moving beyond heuristic scaling toward physically grounded, sustainable intelligence.

Prof. Yang Yang (IEEE Fellow)

Prof. Yunwen Chen

Prof. Hai Zhao

Prof. Fenghua Huang