CV

Jianqiao Zeng (曾健桥)

Shanghai, China · zengjianqiao5@gmail.com · GitHub

Incoming Ph.D. student working on deep learning architectures, representation learning, and AI agents.

Education

• Ph.D. in Computer Science, AI³ Institute, Fudan University, Shanghai Academy of AI for Science, 2026.09 – present (incoming)
  • Advisors: Prof. Hao Xiong (Fudan), Prof. Junchi Yan (SJTU)
• B.E. in Artificial Intelligence, Sun Yat-sen University (中山大学), 2022.09 – 2026.06

Publications

CoEvol-NO: State and Coordinate Co-Evolution with an Error-Driven Predictor-Corrector Paradigm

Jianqiao Zeng, Ruocheng Wang, Yanzhi Liu, Hao Xiong, Junchi Yan. (2026). "CoEvol-NO: State and Coordinate Co-Evolution with an Error-Driven Predictor-Corrector Paradigm." ICML 2026 (Oral).

AMQuestioner: Training Critical Thinking with Question-Driven Interactive Argument Maps in Online Discussion

Qiyu Pan, Jianqiao Zeng, Jien Wang, Junyu Liu, Yihan Qiu, Kangyu Yuan, Zhenhui Peng. (2025). "AMQuestioner: Training Critical Thinking with Question-Driven Interactive Argument Maps in Online Discussion." Proc. ACM Hum. Comput. Interact..

Research Experience

• CoEvol-NO: Co-Evolutionary Neural Operator with Error-Driven Predictor-Corrector Framework First author · Advisors: Prof. Junchi Yan, Prof. Hao Xiong · ICML 2026 (CCF-A) Oral · 2025.10 – 2026.05
  • Paradigm innovation: Identified the structural trade-off between geometric sensitivity and dynamic memory in existing neural operators; proposed Co-Evolution as a third paradigm — latent state and mesh sequence evolve jointly through bidirectional updates.
  • Theoretical unification: Proved that residual connections and direct substitution are first-order approximations of a single Error-Driven Correction framework under different loss assumptions; introduced exact Jacobian gradient to upgrade updates to full optimization. Formalized the symmetry-induced stability advantage and condition-number cost.
  • Empirical validation: Achieved SOTA on 5 standard PDE benchmarks and 2 industrial tasks. 40-layer deep extension shows Navier-Stokes error decreases monotonically (48% drop). PC framework transfers consistently to Transolver and LNO. Code
• FreqONet: Frequency-Domain Operator Learning via Dimensional Decomposition First author · Advisor: Prof. Junchi Yan · Manuscript under review · 2025.01 – 2025.07
  • Decoupled-Composed framework: Reformulated high-dimensional PDE solving as a two-stage process — independent 1D coordinate encoding followed by strict composition. Systematically analyzed four composition strategies (Cat / Add / Hadamard / Outer Product); proved tensor product is the unique composition preserving both orthogonality and translation-invariant kernel properties.
  • Universal approximation: Based on Stone-Weierstrass theorem, proved that if each 1D encoding subspace is dense in $C(K_i)$, then the tensor-product span is dense in $C(K_1 \times \cdots \times K_d)$.
  • Hybrid encoding (FreqPE): Decomposed encoding space into fixed-orthogonal basis (high-frequency structure prior) plus learnable basis (low-frequency data-driven) with a complementary loss enforcing orthogonality.
• AMQuestioner: Training Critical Thinking with Question-Driven Interactive Argument Maps in Online Discussion Co-first author · Advisor: Prof. Zhenhui Peng · CSCW 2025 · 2023.11 – 2024.10
  • Designed and developed a critical-thinking training system combining argument maps with LLMs.
  • Frontend (Vue.js) for mind-map visualization and question-driven interaction; backend (Flask RESTful) for LLM role-play prompt design and adaptive opinion exploration.
  • Led user-experiment design, execution, and data analysis; wrote the core paper sections.
• Learning Agent: Problem-DAG-Driven Personal Knowledge Tracking and Adaptive Learning Independent project · 2026.04 – present
  • Core insight: “Mastering a domain = answering its core question set.” Models knowledge as a DAG (multi-parent + cross-domain links); learning progress measured by question-decomposition depth, not reading volume.
  • Methodology & feedback loop: Designed 5 question decomposition methods (prerequisite / component / contrast / causal chain / assumption clarification); AI generates questions only, not answers. CHECK-DECIDE mechanism propagates answer updates via the DAG, marking downstream nodes as STALE for consistency cascade.
  • Engineering: React + Vite + Tailwind static frontend with infinite SVG canvas, three views (radial / text tree / focus navigator), global knowledge graph, statistics dashboard. 9 Claude Code skills + Python CLI for node lifecycle management.

Selected Open-Source Projects

• Skill-Manager — Skill lifecycle management tool (registration, call counts, user feedback collection, A/B evaluation).
• CoEvol-NO — Co-evolutionary neural operator framework. GitHub
• Learning Agent — Graph-driven dialogue-based learning system.

Skills

• Programming: Python (expert), C/C++, CUDA, Bash, LaTeX
• Deep Learning: PyTorch (expert), JAX, Hugging Face Transformers
• Tools: Claude Code, Git, Linux, Docker
• Theory: Matrix theory, functional analysis, convex optimization
• Reproductions: PINN, FNO, CNO, DeepONet, DDPM, DiT

Professional Service

• (To be added — conference reviewing, mentoring, open-source contributions)