CV
Curriculum vitae — education, experience, and publications.
Contact Information
| Name | Tingzhu Bi |
| Professional Title | PhD Candidate, Peking University |
| bitingzhu@stu.pku.edu.cn | |
| Location | Beijing, China |
| Website |
Professional Summary
PhD candidate at Peking University working on accountable LLM agents for reasoning and complex-system diagnosis, dynamic causal discovery, and AIOps. Research grounded in real hyperscale production data through a long-running collaboration with ByteDance. Expected graduation: 2027.
Experience
-
2021.12 - 2023.03 Beijing, China
Research Intern, AIOps (Data-SYS-STE)
ByteDance
Research on metric causal-relationship discovery and dynamic fault-propagation networks over ByteDance’s hyperscale (million-host) data centers, supporting the intelligent-operations platform. Led to FaultInsight (KDD’24).
- Deep causal fault diagnosis over heterogeneous host-level metrics
- Industry collaboration on real production incident data
-
2020.05 - 2020.08 Beijing, China
iOS Developer Intern, Lark iOS (Enterprise Service – Lark – iOS)
ByteDance
Contributed to the refactor of the Lark Feed page, focusing on first-screen load time and scrolling smoothness. My algorithmic improvement passed performance-evaluation acceptance and was merged into the production release.
Education
-
2023 - 2027 PhD in Software Engineering
Peking University
School of Software & Microelectronics
-
2020 - 2023 Master in Software Engineering
Peking University
School of Software & Microelectronics
-
2015 - 2019 Bachelor in Computer Science
Xidian University
School of Computer Science and Technology
Publications
-
2026 Not Until the Evidence Says So: Teaching LLM Investigators When to Close a Case
Preprint
Studies when an LLM investigator should close a case versus leave it open and name what is missing; introduces Nautil (731 audited cases across aviation, rail, maritime, chemical-safety, vehicle-defect reports and production server incidents) and a three-part closure evaluation; fine-tuning plus closure-only RL makes a 9B model’s closures follow the evidence.
-
2026 JustDiag: A Diagnostic Justification Engine for Accountable Root Cause Analysis
arXiv preprint 2606.19407
A diagnostic justification engine that maintains an explicit, auditable process state over evidence, hypotheses, and conflicts; evaluated on 66 real-world incidents with a two-layer outcome/process protocol.
-
2026 CAVIAR: Disentangling Root Causes with an ICA-based VAE for Large-Scale Microservice Systems
ACM SIGKDD (KDD)
ICA-based variational autoencoder that disentangles independent causal factors to improve root-cause identification in large-scale microservice systems. (Co-author.)
-
2025 UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems
Advances in Neural Information Processing Systems (NeurIPS)
Uncoupler-Recoupler representation disentanglement with temporal-perturbation analysis to recover time-resolved dynamic causal graphs; outperforms SOTA on static benchmarks while capturing evolving causality.
-
2024 G-Cause: Parameter-free Global Diagnosis for Hyperscale Web Service Infrastructures
IEEE International Conference on Web Services (ICWS)
Parameter-free global fault diagnosis across hyperscale web-service infrastructures. (Co-author.)
-
2024 FaultInsight: Interpreting Hyperscale Data Center Host Faults
ACM SIGKDD (KDD)
Interpretable deep causal framework for host-fault diagnosis over heterogeneous host-level metrics; multi-perspective diagnostic insights validated on real ByteDance production incidents.
-
2023 HEAL: Performance Troubleshooting Deep inside Data Center Hosts
ACM SIGMETRICS / POMACS 7(3)
Performance troubleshooting deep inside data-center hosts via heterogeneous host-level signals. (Co-author.)
-
2022 VECROsim: A Versatile Metric-oriented Microservice Fault Simulation System
IEEE ISSRE (Tools and Artifact Track)
Open, customizable microservice fault-simulation system and benchmark dataset providing a standardized data foundation for fault-diagnosis and causal research.