Beihang NLP researcher linking LLM unlearning and interpretability through geometric representation analysis
2
papers here
2
reviewed by Evolve
1
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—
No publication counts, venue history, or yearly output totals are available in the snapshot, so output is unscored. Once linked to a verified publication record, this axis would measure cumulative paper volume, first-author share, and output consistency over time.
Impact—
No citation counts, h-index, or per-paper citation data exist in the snapshot, so impact is unscored. With citation records linked, it would measure aggregate citations, highly cited papers, and influence within the unlearning and interpretability communities.
Focus0/100
The field distribution splits exactly evenly — 0.5 to natural language processing (LLM interpretability) and 0.5 to natural language processing (LLM unlearning and privacy) — so no single field dominates and the radar focus score is recorded as 0. Since both fields sit inside NLP, the profile reads as a coherent geometry-plus-safety agenda rather than scattered topic-hopping.
Collaboration55/100
The snapshot records 15 distinct coauthors, with 12 recurring collaborators including Shiying Duan, Zhichao Yang, and Yuanze Hu; the radar collaboration score is 55. This points to a mid-sized, lab-anchored network typical of an early-career researcher publishing with a Beihang group cluster rather than a large distributed web of one-off partners.
Activity—
The snapshot provides no publication-date series beyond the two 2026 arXiv preprints, so activity is unscored. Once yearly submission records are linked, it would measure recent publishing cadence, gaps between outputs, and sustained engagement with the two subfields.
Reproducibility55/100
Evolve assessed 2 papers with a mean reproducibility score of 55, both rated medium readiness. Reviews flag concrete reproduction risks: the Cascade paper's near-null hyperbolic ablation and Qwen extraction gap, and a possible cluster confound behind the atlas's δ≈1 plateau.
Research summary
Qingchen Yu is a natural language processing researcher at Beihang University whose reviewed work sits at the junction of LLM unlearning and privacy and LLM interpretability, unified by a geometric view of model internals. As first author of 'Cascade: Hierarchical Recoverability Control for Large Language Model Unlearning' (2026, arXiv), Yu proposes a three-tier forgetting-control scheme that reduces recoverable knowledge; Evolve's review notes strong results on TOFU and MUSE, while flagging an extraction gap on Qwen and a near-null hyperbolic ablation that weaken the paper's conclusions. As a middle author on 'A Hyperbolicity Atlas of Large Language Model Hidden States' (2026, arXiv), Yu contributed to the first systematic Gromov-hyperbolicity atlas of LLM prompt hidden states, covering 10 models and 4 datasets, with solid empirics whose δ≈1 plateau may reflect a cluster confound. The bibliometric snapshot records 15 distinct coauthors, a field distribution split evenly at 0.5 between the two subfields, and a reproducibility mean of 55 across the 2 reviewed papers, both rated medium readiness. Radar scores are available for collaboration (55), reproducibility (55), and focus (0); output, impact, and activity axes lack underlying data and are left unscored. The profile is name-resolved from public bibliographic data.
Collaboration pattern
The snapshot lists 15 distinct coauthors, 12 of them recurring, including Shiying Duan, Zhichao Yang, Yuanze Hu, Gen Li, Xinyu Wang, Ye Qiu, Zeming Liu, Guangxu Chen, Zhaoxin Fan, Xiaodong Li, Yuhua Wang, and Zhiyu Li; the radar collaboration score is 55. The pattern suggests lab-anchored collaboration at Beihang University, pairing the first-author unlearning work with co-authored interpretability studies, but without affiliation data per coauthor the network's structure cannot be confirmed.
Research keywords
Highlights
First-author Cascade framework for hierarchical recoverability control in LLM unlearning
Co-authored first systematic Gromov-hyperbolicity atlas of LLM hidden states
A geometry-driven thread connecting unlearning and interpretability
Across both reviewed papers, hyperbolic and tree-like geometric representations recur — as the representation basis in Cascade's recoverability control and as the measurement object in the hyperbolicity atlas — giving the profile a distinctive geometry-driven methodology within NLP safety research; the near-null hyperbolic ablation in Cascade suggests this thread still needs stronger evidence.
Timeline
- 2026Published first-author arXiv preprint Cascade: Hierarchical Recoverability Control for Large Language Model Unlearning in the LLM unlearning and privacy field.
- 2026Co-authored as a middle author the arXiv preprint A Hyperbolicity Atlas of Large Language Model Hidden States in the LLM interpretability field.
- 2026Both 2026 preprints entered the Evolve review catalog with medium reproducibility readiness, at a mean score of 55 across the 2 papers.
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 and co-author counts come from a public bibliometric snapshot; no ORCID or explicit author disambiguation is provided, so this profile is name-resolved from public bibliographic data.
- The output, impact, and activity axes have no underlying data in this snapshot (no citation counts, publication totals, or yearly cadence); they are left unscored rather than estimated.
- Only the 2 papers reviewed by Evolve are in scope; this profile reflects that small reviewed sample, not the researcher's full publication record.
- Both reviewed papers are 2026 arXiv preprints, and the reproducibility mean of 55 is computed from those 2 papers only; no peer-review or venue acceptance data is included.