Beihang NLP researcher linking LLM interpretability and unlearning via hyperbolic representation geometry
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Insufficient data: this axis measures publication volume and productivity trends, but no verified counts of papers, venues, or per-year output are available in the current snapshot, so no score is computed and none is estimated here.
Impact—
Insufficient data: this axis measures citation-based influence — total citations, h-index, and uptake by later work — none of which are available for this profile yet, so impact is left unscored rather than approximated from thin evidence.
Focus0/100
The radar records a focus score of 0/100. The observed field distribution is split exactly evenly — 0.5 natural language processing (LLM interpretability) and 0.5 natural language processing (LLM unlearning and privacy) — so no single subfield dominates. Both cataloged papers nonetheless share one methodological thread, hyperbolic and tree-like representation geometry applied to LLM hidden states, the closest signal of a stable research identity in the current data.
Collaboration55/100
Collaboration scores 55/100, based on 15 distinct coauthors in the observed snapshot. Twelve recurring partners are recorded, including Qingchen Yu, Zhichao Yang, Yuanze Hu, Gen Li, Xinyu Wang, Ye Qiu, Zeming Liu, Guangxu Chen, Zhaoxin Fan, Xiaodong Li, Yuhua Wang, and Zhiyu Li. Both cataloged papers place Duan in a middle, non-corresponding author position, consistent with contributing inside multi-author teams rather than directing them.
Activity—
Insufficient data: this axis measures recent research tempo — papers per year and continuity of output over time — but the snapshot covers only two reviewed 2026 papers, so no activity trend or score can be established yet.
Reproducibility55/100
Reproducibility scores 55/100, computed as a mean of 55 over 2 reviewed papers. Both carry a medium reproducibility readiness rating in the Evolve catalog: the hyperbolicity atlas reports experiments across 10 models and 4 datasets, while Cascade is evaluated on TOFU and MUSE. Medium readiness signals that experimental detail is partially sufficient for independent replication but not fully turnkey.
Research summary
Shiying Duan is a natural language processing researcher at Beihang University whose reviewed work splits evenly between two subfields — LLM interpretability and LLM unlearning and privacy, each 0.5 of the observed field distribution — with hyperbolic and tree-like representation geometry as the shared methodological thread. In 2026 Duan co-authored 'A Hyperbolicity Atlas of Large Language Model Hidden States', which Evolve reviewers describe as the first systematic Gromov-hyperbolicity atlas of LLM prompt hidden states, covering 10 models and 4 datasets with solid empirics, though the δ≈1 plateau may be a cluster confound. The same year's 'Cascade: Hierarchical Recoverability Control for Large Language Model Unlearning' introduces three-tier control that reduces recoverable knowledge and performs strongly on TOFU and MUSE, while a Qwen extraction gap and a near-null hyperbolic ablation weaken some conclusions. Both papers are middle-author, non-corresponding contributions with medium reproducibility readiness, giving a mean reproducibility score of 55 across 2 papers, and the observed network spans 15 distinct coauthors. Output, impact, and activity metrics are not yet available, so this should be read as an early-stage profile.
Collaboration pattern
The observed collaboration network contains 15 distinct coauthors, with a collaboration score of 55. Twelve recurring partners are recorded, including Qingchen Yu, Zhichao Yang, Yuanze Hu, Gen Li, Xinyu Wang, Ye Qiu, Zeming Liu, Guangxu Chen, Zhaoxin Fan, Xiaodong Li, Yuhua Wang, and Zhiyu Li. Both cataloged papers are middle-author, non-corresponding contributions, suggesting a contributor role within multi-author teams; co-author affiliations beyond Beihang University are not recorded in the data.
Research keywords
Highlights
First systematic hyperbolicity atlas of LLM hidden states
Hierarchical recoverability control for LLM unlearning
Timeline
- 2026Co-authored 'Cascade: Hierarchical Recoverability Control for Large Language Model Unlearning' (arXiv) in a middle author position; Evolve rated its TOFU/MUSE results strong while noting a Qwen extraction gap.
- 2026Co-authored 'A Hyperbolicity Atlas of Large Language Model Hidden States' (arXiv), credited as the first systematic Gromov-hyperbolicity atlas of LLM prompt hidden states across 10 models and 4 datasets.
- 2026Both papers entered the Evolve review catalog with medium reproducibility readiness, yielding a mean reproducibility score of 55 over 2 papers and an even 0.5/0.5 field split between LLM interpretability and LLM unlearning.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric scores come from a narrow OpenAlex-style snapshot; the output, impact, and activity axes lack sufficient data and are deliberately left unscored.
- Paper-level analysis rests on only two reviewed papers, both 2026 arXiv preprints; no peer-reviewed venue record is included.
- No ORCID or country is provided, and 'Shiying Duan' may not be a unique name; this profile is name-resolved from public bibliographic data tied to the Beihang University affiliation.
- The reproducibility score (55) is based on just 2 papers rated medium, and the focus score of 0 is the recorded radar value, not an independent audit.