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. I study LLM reasoning and agentic tool use, with research in evaluation, RL post-training, and mechanistic interpretability.

My two representative papers are Peak-Then-Collapse, published in the EMNLP 2026 Main Conference, and KinshipQA, published in Findings of EMNLP 2026. The first investigates GRPO training collapse and interface feedback in knowledge-graph tool use. The second provides a procedurally generated benchmark across seven kinship systems and up to six reasoning hops, with controlled evaluation of cultural-rule application.

Two further manuscripts are under review: Tool-Call Dependency Structure is Linearly Decodable, on dependency representations in agent residual streams, and From Target Selection to Condition Reuse, on the limits of residual interventions. Both are available as arXiv preprints, with code and an archived demo for the latter.

Earlier work includes KGEIR, a framework combining knowledge graphs with iterative LLM reasoning. I am exploring research and applied-science roles in LLM/Agent evaluation and post-training in the UK and China.

CV · GitHub · LinkedIn · Google Scholar

News

Oct 27, 2026	Publication update: From Target Selection to Condition Reuse is available on arXiv. The official EMNLP 2026 Main and Findings versions of Peak-Then-Collapse and KinshipQA are now available in the ACL Anthology.
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.
May 25, 2026	New paper Peak-Then-Collapse and the Four Interface Channels of Knowledge-Graph Tool Use available on arXiv.

Selected Publications

EMNLP Main
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, Oct 2026

Abs arXiv Bib HTML PDF 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}, address = {Budapest, Hungary}, year = {2026}, month = oct, publisher = {Association for Computational Linguistics}, pages = {10214--10230}, url = {https://aclanthology.org/2026.emnlp-main.443/}, }

Findings

Kinship Data Benchmark for Multi-hop Reasoning

Tianda Sun and Dimitar Kazakov

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

Abs arXiv Bib HTML PDF Code

@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},
  month = oct,
  publisher = {Association for Computational Linguistics},
  pages = {5767--5784},
  url = {https://aclanthology.org/2026.findings-emnlp.274/},
}