CV
Contact Information
| Name | Tianda Sun |
| Professional Title | PhD Candidate in Computer Science |
| tianda.sun@york.ac.uk | |
| Website | https://tiandasun.github.io/ |
Professional Summary
Final-year Computer Science PhD working on LLM reasoning and agentic tool use, with research in evaluation, RL post-training, and mechanistic interpretability. Two first-author papers published at EMNLP 2026: Main Conference and Findings.
Experience
-
2022 - present York, United Kingdom
PhD Research — LLM Reasoning and Agentic Tool Use
University of York
Doctoral research supervised by Dr. Dimitar Kazakov.
- Built a single-GPU GRPO post-training pipeline with verl and vLLM; used multi-seed experiments and reward controls to diagnose training collapse and investigate interface feedback. Published in EMNLP 2026 Main Conference.
- Created KinshipQA across seven kinship systems and up to six reasoning hops, with an RDF/OWL and SPARQL generation pipeline and evaluation of six LLMs. Published in Findings of EMNLP 2026.
- Studied tool-call dependency representations with linear probes, controls, and activation patching across two model families.
- Audited residual interventions on target selection, condition reuse, independent eligibility judgments, and complete actions; released reproducible research artifacts.
-
2022 - 2025 York, United Kingdom
Graduate Teaching Assistant
University of York
- DEEP: Deep Learning and Evolution of AI Paradigms
- EVAC: Evaluating and Validating AI Claims
Education
-
2022 - present York, United Kingdom
PhD
University of York
Computer Science
- Thesis: Knowledge Graph-Guided Reasoning in Large Language Models: Towards Interpretable Question Answering
- Supervisor: Dr. Dimitar Kazakov
-
2021 - 2022 York, United Kingdom
MRes
University of York
Computer Science
- Thesis: Relation Extraction in Financial Reports using Distant Supervision with BERT and PCNN
-
2017 - 2018 Liverpool, United Kingdom
MSc
University of Liverpool
Advanced Computer Science
-
2011 - 2015 Tianjin, China
BSc
Tianjin Polytechnic University
Software Engineering
Publications
-
2026 Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use
EMNLP 2026 Main Conference, pp. 10214–10230
Tianda Sun, Dimitar Kazakov
-
2026 Kinship Data Benchmark for Multi-hop Reasoning (KinshipQA)
Findings of EMNLP 2026, pp. 5767–5784
Tianda Sun, Dimitar Kazakov
-
2025 KGEIR: Knowledge Graph-Enhanced Iterative Reasoning for Multi-Hop Question Answering
RANLP 2025 Workshop on Robust and Low-Resource Large Language Models (R2LM), Varna, Bulgaria
Tianda Sun, Dimitar Kazakov
-
2025 TOGGLE Delineates Fate and Function within Individual Cell Types via Single Cell Transcriptomics
bioRxiv preprint
Jiawei Chen, Zhihao Chen, Tianda Sun, Kun Liu, Enming Jiang, Yongjie Nong, Tao Yuan, Chun-Chun Dai, Yao Yan, et al.
-
2024 A Hybrid Question Answering Model with Ontological Integration for Environmental Information
DAO-XAI 2024 Workshop at ECAI, Santiago de Compostela, Spain
Tianda Sun, Jamie Carr, Dimitar Kazakov
-
2022 NLP Analysis of COVID-19 Radiology Reports in Indonesian using IndoBERT
IEEE International Biomedical Instrumentation and Technology Conference (IBIOMED)
Nunung Nurul Qomariyah, Tianda Sun, Dimitar Kazakov
Preprints and Manuscripts Under Review
- Tool-Call Dependency Structure is Linearly Decodable in LLM Agent Residual Streams — Tianda Sun and Dimitar Kazakov (2026). Public arXiv preprint; under review. Linear probing and activation patching of tool-call dependency representations. Code.
- From Target Selection to Condition Reuse: Auditing Residual Interventions in Tool-Use Models — Tianda Sun and Dimitar Kazakov (2026). Public arXiv preprint (arXiv:2610.31686); under review. Controlled evaluation of residual interventions in Qwen3.5-27B across 32 fresh worlds. Code and archived demo.
Skills
Programming (Expert): Python, Java, SQL, Bash, C++
ML/DL Frameworks (Advanced): PyTorch, Hugging Face Transformers, scikit-learn, GRPO, verl, vLLM, LangChain
Evaluation & Interpretability: Benchmark design, multi-seed and cross-model evaluation, linear probing, activation patching
NLP & Knowledge Graphs (Advanced): SPARQL, RDF/OWL, Neo4j, NLTK, spaCy
Infrastructure (Advanced): Linux, Git, Docker, HPC clusters