Tianda Sun

Artificial Intelligence Group

Department of Computer Science

University of York

York, United Kingdom

I am a final-year PhD candidate in Computer Science at the University of York, supervised by Dr. Dimitar Kazakov. My research focuses on LLM reasoning, knowledge graph integration, and benchmark development for natural language understanding.

I created KinshipQA, a contamination-proof benchmark for evaluating multi-hop reasoning across culturally diverse contexts, and developed the KGEIR framework that integrates knowledge graphs with iterative LLM reasoning.

My current work extends to agentic tool use and mechanistic interpretability of LLM agents: Peak-Then-Collapse studies reinforcement-learning post-training for knowledge-graph tool use and the role of interface feedback, while Tool-Call Dependency Structure is Linearly Decodable probes how agents internally represent tool-call dependencies in their residual streams.

Research Interests: Large Language Model Reasoning, Multi-hop Question Answering, Knowledge Graph Construction & Reasoning, Agentic Tool Use & Reinforcement Learning, Mechanistic Interpretability, Retrieval-Augmented Generation (RAG), Benchmark Development & Evaluation, Cross-cultural NLP

News

Aug 26, 2026	KinshipQA paper accepted to Findings of EMNLP 2026 in Budapest, Hungary.
Aug 26, 2026	Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use accepted to the EMNLP 2026 Main Conference in Budapest, Hungary.
May 25, 2026	New paper Tool-Call Dependency Structure is Linearly Decodable in LLM Agent Residual Streams available on arXiv; under review in the ARR August 2026 cycle, targeting EACL 2027.
May 25, 2026	New paper Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use available on arXiv.
Jan 07, 2026	Paper on KinshipQA benchmark available on arXiv.

Selected Publications

EMNLP
Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use

Tianda Sun and Dimitar Kazakov

In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026

Abs arXiv Bib Code

We test the standard RLVR tool-use recipe (GRPO on Qwen2.5-7B-Instruct) on a deliberately minimal knowledge-graph tool API of four Freebase navigation verbs over Complex WebQuestions. The policy’s tool-grounded answer rate climbs from 3.8% to 9.6% before collapsing to 0% within a single 50-step window, a peak-then-collapse pattern replicated across four seeds and seven reward designs. We trace the failure to sparse interface feedback and show that one-iteration self-distillation reaches 40.0% exact-match accuracy, with the performance ceiling appearing interface-bound across the 7B-14B range.
@inproceedings{sun2026peakcollapse, title = {Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use}, author = {Sun, Tianda and Kazakov, Dimitar}, booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)}, address = {Budapest, Hungary}, year = {2026}, }
arXiv
Tool-Call Dependency Structure is Linearly Decodable in LLM Agent Residual Streams

Tianda Sun and Dimitar Kazakov

arXiv preprint arXiv:2605.25310, 2026

Abs arXiv Bib Code

Tool-using LLM agents produce trajectories whose calls form a directed dependency graph, where earlier tool outputs supply arguments to later calls. Using a low-capacity edge probe on the residual stream of Qwen3-32B, we decode this dependency graph well above both a Hewitt-Liang random-label control and a positional baseline, with counterfactual tests indicating the signal tracks abstract topology rather than identifier values. To our knowledge this is the first structural probe of an LLM agent’s runtime tool-call dependency graph, replicating across three multi-hop benchmarks and two model families.
@article{sun2026toolcall, title = {Tool-Call Dependency Structure is Linearly Decodable in {LLM} Agent Residual Streams}, author = {Sun, Tianda and Kazakov, Dimitar}, journal = {arXiv preprint arXiv:2605.25310}, year = {2026}, }
EMNLP
Kinship Data Benchmark for Multi-hop Reasoning

Tianda Sun and Dimitar Kazakov

In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

Abs arXiv Bib Code

We introduce KinshipQA, a contamination-proof benchmark evaluating LLM multi-hop reasoning across 7 culturally diverse kinship systems. We evaluate 6 state-of-the-art LLMs on 3,134 questions with controlled reasoning complexity (1-4 hops), revealing significant performance degradation as reasoning depth increases.
@inproceedings{sun2026kinshipqa, title = {Kinship Data Benchmark for Multi-hop Reasoning}, author = {Sun, Tianda and Kazakov, Dimitar}, booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026}, address = {Budapest, Hungary}, year = {2026}, }
RANLP
KGEIR: Knowledge Graph-Enhanced Iterative Reasoning for Multi-Hop Question Answering

Tianda Sun and Dimitar Kazakov

In Proceedings of the RANLP 2025 Workshop on Robust and Low-Resource Large Language Models (R2LM), 2025

Abs Bib HTML Code

We propose KGEIR, a framework combining LLM reasoning with dynamic knowledge graph construction. KGEIR achieves competitive and superior performance on HotpotQA, 2WikiMultiHopQA, and MuSiQue benchmarks compared to state-of-the-art RAG methods, with a 4.5% F1 improvement over Tree-of-Thought baseline.
@inproceedings{sun2025kgeir, title = {KGEIR: Knowledge Graph-Enhanced Iterative Reasoning for Multi-Hop Question Answering}, author = {Sun, Tianda and Kazakov, Dimitar}, booktitle = {Proceedings of the RANLP 2025 Workshop on Robust and Low-Resource Large Language Models (R2LM)}, year = {2025}, address = {Varna, Bulgaria}, }
ECAI
A Hybrid Question Answering Model with Ontological Integration for Environmental Information

Tianda Sun, Jamie Carr, and Dimitar Kazakov

In Proceedings of DAO-XAI 2024 Workshop at ECAI, 2024

Abs Bib PDF

We present a hybrid question answering model that integrates ontological knowledge for environmental information retrieval, combining structured knowledge representation with neural language understanding.
@inproceedings{sun2024hybrid, title = {A Hybrid Question Answering Model with Ontological Integration for Environmental Information}, author = {Sun, Tianda and Carr, Jamie and Kazakov, Dimitar}, booktitle = {Proceedings of DAO-XAI 2024 Workshop at ECAI}, year = {2024}, address = {Santiago de Compostela, Spain}, }