Interpretability-focused researcher spanning vision-language evidence localization and LLM reranking calibration
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-19) 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—
Output is not yet assessable: the snapshot records h-index 0, i10-index 0 and a single 2026 work — too thin to score. This axis would measure total publication volume, authorship positions and year-by-year productivity.
Impact—
Impact is not yet assessable: no citation accumulation is visible (h-index 0, i10-index 0), and both catalog papers are recent 2026 preprints. This axis would measure citation counts, yearly citation trends and uptake by later work.
Focus0/100
Focus scores 0 because the record is perfectly split: the OpenAlex field distribution assigns 0.5 to computer vision (vision-language models) and 0.5 to natural language processing / information retrieval, while the three logged topics — topic modeling, generative adversarial networks and image synthesis, and multimodal machine learning applications — each appear once. Across the two catalog papers, the work straddles the vision-language and retrieval sides rather than one niche.
Collaboration46/100
Collaboration scores 46, built on 9 distinct coauthors in the recent window. The catalog papers show team-based writing: first author on ProtoLIP and a middle-position author on the reranking calibration study. Recurring names include Yongbo Chen, Zhengming Ding, Petros Karypis and Julian McAuley, pointing to involvement in larger NLP and vision teams rather than solo or two-person projects.
Activity20/100
Activity scores 20: the counts-by-year record shows a single work in 2026 and recent three-year works total 1, so the observed publishing rate is low. Both catalog entries are 2026 arXiv preprints, meaning the visible record begins this year; activity may be understated if earlier work is not linked to this name-resolved identity.
Reproducibility55/100
Reproducibility scores 55, computed from 2 papers with a mean readiness of 55. Both Evolve-reviewed papers — ProtoLIP and the attention-based reranking study — are rated medium readiness: methods such as decomposable prototypes and training-free interpolated nulls are described concretely, but the verdicts flag practical frictions including retrieval drops, backbone-dependent transfer, and RankGPT still leading on FollowIR.
Research summary
Yan Zhu is a researcher whose visible record centers on making large AI models more transparent, with work split evenly between computer vision (vision-language models) and natural language processing / information retrieval — the OpenAlex field distribution assigns 0.5 to each side. The Evolve catalog includes two 2026 arXiv papers. The first, ProtoLIP, lists Yan Zhu as first author: it pushes evidence disentanglement in vision-language models from sentence level to object level without spatial supervision, using decomposable prototypes; Evolve's verdict notes retrieval drops and backbone-dependent transfer. The second, a middle-author contribution, examines attention-based reranking and shows that null-query calibration breaks under per-query instructions, with training-free interpolated nulls as the fix — though RankGPT still wins on FollowIR. Bibliometric signals are early-stage: h-index and i10-index are both 0, one work is recorded for 2026, recent three-year works number 1, and 9 distinct coauthors appear in the window. Reproducibility is moderate, with a mean readiness of 55 across 2 reviewed papers, both rated medium. This profile is name-resolved from OpenAlex public bibliographic data; institutional affiliation is not available in the provided inputs.
Collaboration pattern
Yan Zhu's visible collaborations are team-based, with 9 distinct coauthors in the recent window. On the catalog papers, Yan Zhu is first author on ProtoLIP and a middle author on the reranking calibration study. Recurring names — Yongbo Chen, Zhengming Ding, Rebecca Faust, Petros Karypis, Hossein Rajaby Faghihi, Noveen Sachdeva and Julian McAuley — indicate ties to both NLP and vision groups; the collaboration axis scores 46, mid-range.
Research keywords
Highlights
First-author ProtoLIP advances object-level evidence localization in VLMs
Diagnoses and repairs null-query calibration in LLM rerankers
A 2026 portfolio spanning VLM interpretability and retrieval calibration
Both catalog papers appeared in 2026, split evenly across vision-language models and NLP / information retrieval, and each is rated medium reproducibility readiness with a mean of 55.
Timeline
- 2026First-author arXiv preprint ProtoLIP: From Sentence-Level to Object-Level Evidence Disentanglement appears, targeting object-level localization in vision-language models without spatial supervision.
- 2026Co-authors (middle position) the arXiv study 'How Calibration Content Shapes Attention-Based Reranking' in natural language processing / information retrieval.
- 2026OpenAlex snapshot records 1 work in 2026 with 9 distinct coauthors, while h-index and i10-index both stand at 0 — a single-year, early-stage record.
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.
- Common-name caution: 'Yan Zhu' is name-resolved from OpenAlex public bibliographic data; no ORCID or institutional affiliation was available to confirm identity.
- All bibliometrics come from an OpenAlex snapshot; citation coverage may be incomplete for very recent 2026 arXiv preprints.
- Output and impact axes are unscored because the snapshot shows h-index 0, i10-index 0 and only one recorded work.
- Both catalog papers are arXiv preprints; peer-reviewed venue status is not reflected in the provided data.