First-author work on fair VLM answer decoding and scoped low-rank knowledge editing for large language models
2
papers here
1
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-28) 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 output metrics are available in the current snapshot — publication counts, venue spread, and author-position trends could not be computed from the input. This axis would measure the volume and trajectory of published work over time.
Impact—
Citation-based influence data — citations, h-index, citing works — is absent from the input, so impact is unscored. This axis would measure how widely the cataloged papers and any other work are taken up by the research community.
Focus100/100
Focus scores 100 because the cataloged field distribution is a single area — machine learning / NLP, with the knowledge-editing paper as the one classified record. Both 2026 papers, VLM answer decoding and LLM knowledge editing, sit inside one coherent model-internals-and-evaluation agenda with no scatter across unrelated fields.
Collaboration28/100
The collaboration footprint is compact: 3 distinct coauthors across the cataloged work, with Jianfeng Xu, Hu Xu, and Siyuan Qiu as the recurring partners. The score of 28 reflects a small-team, first-author-driven pattern — both 2026 papers are first-authored and non-corresponding — rather than a broad multi-group network.
Activity—
Publishing cadence cannot be assessed: the snapshot holds only two papers, both dated 2026, with no multi-year history. This axis would measure publication rate and continuity across successive years once a longer record exists.
Reproducibility55/100
Reproducibility averages 55 across one assessed paper (reproPapers: 1), and both cataloged papers carry medium reproducibility readiness. The decoding study's claims rest on strong but narrow evidence without significance testing, while ALOE's scope results (effectiveness 0.955–0.999, localization 0.981–1) sit alongside possibly undertuned baselines — independent replication is advisable before reuse.
Research summary
Zeyan Li is an early-career researcher working at the intersection of large language models and their evaluation, profiled here from a focused Evolve catalog of two 2026 arXiv papers on which Li is first author. The work splits into two tightly related threads. The first concerns how we measure multimodal models: "Reasoning Instructions Can Break Answer Decoding in Vision–Language Models" shows that scoring multiple-choice answers after a chain-of-thought request — reading label logits behind a CoT prefix — understates VLM accuracy by 35.28 points on ScienceQA, and proposes linear probes and direct logit analysis as fairer decoding; Evolve rates the evidence strong but coverage narrow and without significance testing. The second thread targets model internals: "ALOE: Semantically Addressed Low-Rank Operators for Knowledge Editing" introduces gated low-rank addresses that fix edit scope, with effectiveness of 0.955–0.999 and localization of 0.981–1, though baselines appear undertuned and generalization is weak. Both papers carry medium reproducibility readiness, and the aggregate reproducibility score is 55 from one assessed paper. The radar profile shows extreme focus (100) — the cataloged field distribution is a single NLP/machine-learning area — and a compact collaboration footprint (28) with three distinct coauthors: Jianfeng Xu, Hu Xu, and Siyuan Qiu. Output, impact, and activity remain unscored for lack of citation and volume data, so this should be read as an early-stage, name-resolved profile.
Collaboration pattern
The cataloged record shows a compact, first-author-driven collaboration pattern: 3 distinct coauthors in total, with Jianfeng Xu, Hu Xu, and Siyuan Qiu recurring across the 2026 papers. Both papers list Li as first author and not corresponding, suggesting a hands-on experimental role within a small team rather than leading a large group or multi-institution consortium.
Research keywords
Highlights
CoT-prefix scoring understates VLM accuracy by 35.28 points
Linear probes and logit analysis as fairer decoding
Gated low-rank addresses pin down knowledge-edit scope
Timeline
- 2026Publishes first-author arXiv paper "Reasoning Instructions Can Break Answer Decoding in Vision–Language Models" on evaluation methodology for multimodal models.
- 2026Publishes first-author arXiv paper "ALOE: Semantically Addressed Low-Rank Operators for Knowledge Editing" on scoped knowledge editing for large language models.
- 2026Both papers enter the Evolve catalog with medium reproducibility readiness, produced within a three-person collaborator circle (Jianfeng Xu, Hu Xu, Siyuan Qiu).
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 signals come from a narrow snapshot: only two 2026 arXiv first-author papers reviewed by Evolve; other work by this author may exist outside the catalog.
- No verified institution, country, or ORCID is attached; "Zeyan Li" is a common name, so this profile is name-resolved from public bibliographic data and may not represent a unique individual.
- Output, impact, and activity axes are unscored because citation counts and publication-volume data are absent from the input.
- The reproducibility score of 55 rests on a single assessed paper (reproPapers: 1), so treat all radar scores as early-stage estimates.