cv EN 中
Basics
| Name | Junhao Liu |
| Label | PhD Student |
| liujunhao@pku.edu.cn | |
| Url | https://outerform.site |
| Summary | A PhD student at Peking University, researching eXplainable Artificial Intelligence (XAI). Dedicated to developing explainability technologies that help humans understand, trust, utilize, and control complex AI systems. |
Work
-
2025.07 - Present Research Intern (Project Up), Hunyuan Multimodal Model Team
Tencent
Conducting research and development on HunyuanImage models, focusing on improving model interpretability and controllability, enabling more transparent understanding and precise manipulation of model behavior.
Education
Publications
-
2026 Revitalizing Black-Box Interpretability: Actionable Interpretability for LLMs via Proxy Models
ACL 2026, Main Conference
A proxy-based explanation framework that transfers explanations from budget-friendly models to expensive LLMs, achieving over 90% fidelity at only 11% of the cost.
-
2026 Focus-LIME: Surgical Interpretation of Long-Context Large Language Models via Proxy-Based Neighborhood Selection
IJCAI-ECAI 2026
A coarse-to-fine framework for surgical feature-level interpretation of long-context LLMs via proxy-based neighborhood selection.
-
2025 ReX: A framework for incorporating temporal information in model-agnostic local explanation techniques
AAAI 2025 (Oral, 4.68%)
A general framework that incorporates temporal information into model-agnostic local explanation techniques such as Anchors, LIME, and Kernel SHAP.
-
2026 Guiding LLM-based Loop Invariant Synthesis via Feedback on Local Reasoning Errors
TOPLAS, 2026
A framework that provides constructive feedback to LLMs by formally verifying their reasoning process. The tool LORIS achieved 93.1% success rate on 460 C programs.
-
2026 MAnchors: Memorization-Based Acceleration of Anchors via Rule Reuse and Transformation
ICML 2026
A memorization-based acceleration framework for Anchors that significantly reduces explanation generation time while maintaining fidelity.
Preprints
-
2024 ConLUX: Concept-Based Local Unified Explanations
arXiv:2410.12439
A general framework that extends local model-agnostic techniques to concept-based explanations supporting attributions, sufficient conditions, and counterfactuals.
-
2026 WASD: Locating Critical Neurons as Sufficient Conditions for Explaining and Controlling LLM Behavior
arXiv:2603.18474
A framework that explains model behavior by identifying sufficient neural conditions for token generation.
Awards
- 2023.01.01
Outstanding Research Award
Peking University
- 2022.06.01
Outstanding Graduate of Beijing
Beijing Municipal Education Commission
- 2021.01.01
National Scholarship
Ministry of Education of China
- 2021.01.01
- 2021.01.01
- 2017.07.01
Gold Medal, National Olympiad in Informatics (NOI)
China Computer Federation
Volunteer
-
2020.09 - 2024.12 Teaching Assistant
Peking University
- Introduction to Probabilistic Programming (Graduate Course) - Spring 2024
- Introduction to Discrete Mathematics - Fall 2024
- Programming Practice - Spring 2023
- Introduction to Computation (B) - Fall 2022
- Data Structures and Algorithms Practice - Fall 2020
Projects
- 2023.03 - 2023.06
MTML: A Multi-threaded Language without Data Races and Deadlocks
Designed a multi-threaded programming language based on OCaml, leveraging a type system to statically prevent data races and deadlocks.
- 2023.05 - 2023.06
User-based Collaborative Filtering (Distributed)
Implemented user-based collaborative filtering using Spark and Hadoop, with a comparative study showing Spark's superior efficiency on large-scale workloads.
- 2021.09 - 2021.12
EasyFile: Automated Document Processing Tool
Developed an automated tool for Office and PDF processing, supporting format editing and information extraction.
- 2021.12 - 2022.01
Heuristic EuSolver-based Program Synthesizer
Built a syntax-guided program synthesizer with heuristic rules for CLIA, improving efficiency over standard SMT-based approaches.
- 2021.09 - 2021.11
Java Pointer Analyzer
Implemented a Java pointer analysis tool supporting flow-, context-, and field-sensitive analysis for memory-related bug detection.
Skills
| Programming | |
| C/C++ | |
| Python | |
| Linux | |
| Git | |
| Docker |
| Research | |
| Explainable AI | |
| Machine Learning | |
| Large Language Models |
Languages
| Mandarin | |
| Native speaker |
| English | |
| Fluent (CET-6: 628) |
| Cantonese | |
| Learning |
Interests
| Explainable AI | |
| Model-agnostic explanations | |
| Concept-based explanations | |
| LLM interpretability and control |
| Hobbies | |
| Swimming | |
| Long-distance Running | |
| Sim Racing |