Machine learning researcher working on diffusion-model unlearning and evaluation benchmarks for video-language models
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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—
Publication-volume data — paper counts, venue types, and cumulative production — is not present in the current bibliometric snapshot, so this axis, which would measure total research output over time, cannot yet be scored.
Impact—
Citation counts and other uptake metrics are absent from the input data, so this axis — which would measure the influence of the two 2026 preprints on subsequent research — remains unscored for now.
Focus0/100
The reviewed output splits exactly evenly — 0.5 machine learning (diffusion models, privacy) and 0.5 computer vision (video-language model evaluation) — across the two catalogued 2026 papers, producing a focus score of 0. This reflects a deliberately two-sided profile rather than concentration: one thread builds privacy-preserving diffusion unlearning methods, the other builds evaluation benchmarks for video-language models.
Collaboration51/100
The bibliometric record lists 12 distinct coauthors and a collaboration score of 51, a mid-range level of intensity. Frequent collaborators include Taehyuk Lee, Jinyeong Kim, Youngmin Oh, Dohyeong Kim, Jaehyuk Ryu, Sangwoo Hong, Changbeen Kim, Junwon Chang, Kipyo Kim, Risa Shinoda, and Kuniaki Saito, a pool spanning both the diffusion-unlearning and video-benchmark lines of work.
Activity—
Yearly publication counts and recency signals are missing from the snapshot, so this axis, which would measure publishing cadence and recent research momentum over time, cannot be scored from the current data.
Reproducibility55/100
Reproducibility averages 55 across the two assessed papers (reproPapers = 2), each rated medium reproducibility readiness. Evolve's review flags thin statistics on the diffusion unlearning paper and κ=0.50 annotator agreement on VidOmni-Bench, so both artifacts are documented enough to attempt replication, but verification effort remains moderate rather than trivial.
Research summary
Donghyun Kim is a machine learning researcher whose reviewed work spans two complementary areas: privacy-preserving generative modeling and the evaluation of video-language models. In a 2026 first-author arXiv paper, Kim led LASTING, a method for sequential diffusion data-point unlearning that replaces deletion targets with searched low-loss neighbors, curbing the reappearance of removed content across successive unlearning requests; Evolve's review judges the results strong but the statistics thin, and rates reproducibility readiness medium. In the same year, Kim appeared as last author on VidOmni-Bench, a 500-video benchmark built on spatio-temporal event verification that probes fine-grained video understanding across complexity and duration, exposing hallucination and self-bias in video-LLMs; the design is judged rigorous, with κ=0.50 annotator agreement flagged as a caution. The bibliometric snapshot splits Kim's reviewed output evenly (0.5/0.5) between machine learning (diffusion models, privacy) and computer vision (video-language model evaluation), yielding a focus score of 0. Collaboration scores 51 with 12 distinct coauthors, and reproducibility averages 55 across the two assessed papers. The output, impact, and activity axes lack underlying data and remain unscored.
Collaboration pattern
With 12 distinct coauthors and a collaboration score of 51, Kim operates in a mid-sized network with moderate intensity. The frequent-coauthor list centers on a tightly interlinked cluster — Taehyuk Lee, Jinyeong Kim, Youngmin Oh, Dohyeong Kim, Jaehyuk Ryu, Sangwoo Hong, Changbeen Kim, Junwon Chang, Kipyo Kim — alongside Risa Shinoda and Kuniaki Saito. The data does not map individual coauthors to specific papers, and the list itself contains a duplicate 'Donghyun Kim' entry consistent with name resolution from public bibliographic records.
Research keywords
Highlights
Curbing Reappearance in Sequential Diffusion Unlearning
VidOmni-Bench: Stress-Testing Video-LLMs with 500 Videos
Timeline
- 2026First-author arXiv paper on sequential diffusion data-point unlearning (LASTING) appears in the record.
- 2026Last-author arXiv paper introduces VidOmni-Bench, a 500-video event-verification benchmark for video-language models.
- 2026Both 2026 arXiv preprints are reviewed into the Evolve catalog, each rated medium reproducibility readiness.
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.
- Radar scores are computed from a public bibliometric snapshot; the output, impact, and activity axes lack the underlying citation and volume data and are left unscored.
- Institution, country, and ORCID are not provided, and 'Donghyun Kim' is a common Korean name; this profile is name-resolved from public bibliographic data, and the coauthor list contains an identical name string.
- Both catalog entries are 2026 arXiv preprints; no peer-reviewed venue, citation counts, or publication-history records are included in the input data.
- Evolve review caveats apply to the artifacts themselves: thin statistics on the unlearning paper and κ=0.50 annotator agreement on VidOmni-Bench; both rated medium reproducibility readiness.