Zetao Yang 杨泽涛

AI Scientist Intern @ Apodex · MSc Student @ NTU · Singapore

$$\begin{cases} -\partial_t u + H(x, \nabla u, m) = 0 \\[2pt] \phantom{-}\partial_t m - \mathrm{div}\bigl(m\, \nabla_p H\bigr) = 0 \end{cases}$$
Mean Field Games — the lens that first drew me to populations of agents under shared dynamics. Still the way I think about evaluation, environments, and what we mean by "the policy."
Zetao Yang

Biography

I am an AI Scientist Intern at Apodex (Shanda Group) and an MSc student in Signal Processing & Machine Learning at Nanyang Technological University in Singapore. I build data pipelines for foundation-model pre-training and mid-training, with a focus on training-data quality, traceable releases, and corpus composition. I also work on LLM and agent evaluation, failure attribution, and reproducible benchmarks.

At Apodex, my work covers scientific-document processing, STEM document-quality classification, corpus retrieval and verification, and subject and token composition audits for training data. At ByteDance Seed, I developed tools for agent evaluation and failure diagnosis, including structured analysis of execution traces and an MCP-based post-evaluation agent. At Humanlaya AI, I worked on document-based task synthesis, evidence provenance, rubric auditing, and reproducible evaluation workflows. I enjoy turning ambiguous evaluation problems into explicit assumptions, controlled comparisons, and reusable tools.

I hold a BSc in Mathematics & Applied Mathematics from Beijing Institute of Technology's Strong Foundation Program. I am the sole author of PhysEDA, an arXiv preprint on physics-aware learning for electronic design automation, with research supervision from Prof. Lingfeng Niu, Prof. Xin Shen, and Prof. Qingna Li.

Research Interests

Foundation Model Training Data Pre-training & Mid-training LLM & Agent Evaluation Synthetic Data Quality Reinforcement Learning Linear & Efficient Attention Failure Attribution

Contact

GitHub LinkedIn arXiv Scholar

Research

A guiding question: How can we build training data and evaluations that reliably support the development of foundation models and agents?

My work connects training-data quality, corpus composition, evaluation design, and failure attribution. I use controlled comparisons and reproducible workflows to study data-selection methods and separate model limitations from benchmark defects. My mathematics background also informs an interest in structured physical priors, efficient attention, and environment and reward design for reinforcement learning.

Specific topics:

  1. Foundation-model pre-training and mid-training data
    Processing scientific documents for training, developing STEM document-quality classifiers, and verifying source texts for downstream synthesis. Auditing subject and token composition, with traceable releases and compatible training readers.
  2. Benchmark reliability and controlled evaluation
    Designing matched comparisons and auditing scoring rubrics to distinguish agent failures from conversion defects, answer leakage, and inconsistent evaluation criteria.
  3. Agent failure attribution
    Representing execution traces as decision graphs, developing failure taxonomies, and combining statistical attribution, graph-pattern mining, and targeted perturbations to investigate recurring failures.
  4. Reproducible synthetic tasks
    Building document-based workspaces and tasks with source-evidence mappings, schema-backed generation, temporal-consistency checks, and reproducible outputs. Making data provenance and quality controls explicit throughout the evaluation workflow.
  5. Structured priors for attention
    Incorporating physical priors into linear attention and reward shaping for reinforcement learning. In PhysEDA, separable Manhattan-distance decay reduces attention complexity from O(L²) to O(L). Representative work: PhysEDA (arXiv preprint, 2026).

News

Preprints

PhysEDA: Physics-Aware Learning Framework for Efficient EDA with Manhattan Distance Decay
Zetao Yang
arXiv preprint · 2026 · Sole author
arXiv Code

A unified physics-aware framework injecting Manhattan-distance decay into both the architecture (PSLA — linear attention) and the training stage (PBRS — reward shaping). Evaluated across decoupling-capacitor placement, macro placement, and IR-drop prediction. On 100×100 grids, the PSLA encoder ran 14× faster and used 98.5% less memory than softmax attention in the reported experiments.

Experience

Apodex (Shanda Group)
Sep 2026 – present
AI Scientist Intern · Singapore · Part-time, hybrid
  • Developed scientific-document processing within foundation-model pre-training and mid-training data pipelines, including post-OCR cleanup, LLM annotation, versioned releases, and compatible training readers.
  • Built a STEM document-quality classifier to support mid-training reasoning-data selection, comparing frozen model representations and classification heads through controlled experiments.
  • Built corpus-retrieval and verification workflows for source texts used in downstream synthesis and mid-training, including identity, edition, and deduplication checks.
  • Audited subject and token composition of mid-training corpora and implemented version-checked, reversible metadata corrections with verified readback.
Humanlaya AI
Jun 2026 – Sep 2026
Algorithm Intern · Synthetic Data & Evaluation
  • Built provenance-aware pipelines for document-based tasks and research workspaces, with source mappings and quality controls.
  • Designed paired text-to-file evaluations and audited trajectories and scoring rubrics to distinguish model failures from benchmark defects.
  • Developed schema-backed document generation and paper-to-project-to-question workflows with evidence, temporal-consistency, leakage, and rubric checks; validated outputs through independent reruns.
ByteDance — Seed (Foundation Model Team)
Dec 2025 – Jun 2026
Algorithm Intern · Agent Evaluation
  • Developed an agent-failure taxonomy and confidence-gated trace classifier for interpretable diagnosis.
  • Modelled multistep traces as decision graphs, combining statistical attribution, graph-pattern mining, and targeted perturbations to investigate recurring failures.
  • Deployed an MCP-based post-evaluation agent with structured access to benchmark assets and modular analysis tools in a shared sandbox.
Edgenesis
Jun 2025 – Aug 2025
Edge Large Model Algorithm Intern
Improved prompts and control logic for an IoT protocol-generation agent, addressing structural defects in bidirectional generation. Built Raspberry Pi prototypes for GPS digital-twin synchronisation and speech-enabled retrieval.
TOGEEK
Feb 2025 – May 2025
Algorithm Engineer Intern
RL-based optimisation for cooperative energy storage and photovoltaic generation; defined state/action/reward and implemented the full training pipeline.

Education

Nanyang Technological University, Singapore
Aug 2026 – Feb 2028 (expected)
MSc Student, Signal Processing & Machine Learning · Coursework
Beijing Institute of Technology, China
Sep 2022 – Jun 2026
BSc, Mathematics & Applied Mathematics (Strong Foundation Program)

Awards & Activities