Video understanding & MLLM agent evaluation researcher at Yonsei University; calibration-aware test-time prompt tuning
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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 data is available: the snapshot records no publication counts or venue history for this author. This axis would measure cumulative research production, such as total papers and their spread across years and venues.
Impact—
Impact cannot be scored because the underlying citation data is missing: no citation counts, h-index, or uptake indicators exist here. This axis would measure how widely the work is cited and adopted by the video-understanding and MLLM-evaluation communities.
Focus100/100
Focus registers the maximum score of 100: every paper in the snapshot maps to a single declared field — computer vision / multimodal machine learning — and the field distribution holds exactly one entry, video understanding and MLLM agent evaluation. Both 2026 papers, the AgentVidBench benchmark and the calibration-based test-time prompt tuning study, sit inside vision-language modeling, indicating a tightly concentrated research agenda.
Collaboration46/100
Collaboration registers a mid-range 46, computed from 9 distinct coauthors in the snapshot. The frequent-coauthor list is dominated by close collaborators — Seoyeon An, Minsu Kim, Chanho Lee, Younghan Park (listed twice, implying repeat collaboration), Kangwook Lee, Jungwon Choi, Kibok Lee, and Eunwoo Kim. On AgentVidBench the author position is middle and not corresponding, consistent with a contributing member of a small team rather than a solo effort.
Activity—
Activity data is insufficient: the snapshot spans only 2026 papers and offers no multi-year publication cadence. This axis would measure publishing rhythm over time, such as papers per year and recent momentum.
Reproducibility100/100
Reproducibility scores 100, the maximum, with a reproMean of 100 computed from one assessed paper (reproPapers: 1). That paper, AgentVidBench, carries a 'high' reproducibility-readiness rating; Evolve's verdict notes the judge-validation protocol is solid while flagging benchmark scale and the absence of non-self baselines. With only one paper assessed, this perfect score is encouraging but statistically thin.
Research summary
Hyeonseo Jang is a researcher at Yonsei University whose catalogued work sits in computer vision and multimodal machine learning, with a declared focus on video understanding and the evaluation of MLLM (multimodal large language model) agents. The Evolve catalog contains two 2026 arXiv papers. The first, AgentVidBench, introduces a 100-question multi-hop video question answering benchmark with LLM-judged trajectory scoring; Evolve's verdict credits a solid judge-validation protocol while noting that benchmark scale and a self-only baseline limit its reach, and it rates the paper's reproducibility readiness as high. The second, Rethinking the Test-Time Prompt Tuning Objective from the Perspective of Calibration, moves into adaptation-objective design for vision-language models, though it has not yet received a full Evolve review. Bibliometric signals are narrow but pointed: the field distribution contains a single entry, yielding a perfect focus score of 100, and reproducibility also scores 100, although that mean is computed from just one assessed paper. Collaboration scores a mid-range 46, with 9 distinct coauthors recorded and Younghan Park — listed twice — among frequent partners alongside Seoyeon An, Chanho Lee, Kangwook Lee, and Eunwoo Kim. Output, impact, and activity axes lack underlying data and are left unscored. This profile is name-resolved from public bibliographic data; 'Hyeonseo Jang' is not a rare name, so identity confidence should be treated with care.
Collaboration pattern
Collaboration scores 46 on a base of 9 distinct coauthors. The frequent-coauthor list — Seoyeon An, Minsu Kim, Chanho Lee, Younghan Park (listed twice, signaling repeat collaboration), Kangwook Lee, Jungwon Choi, Kibok Lee, and Eunwoo Kim — points to a stable, small-team pattern. On AgentVidBench the author sits in a middle position without corresponding-author status, consistent with contributing within a group rather than leading solo work.
Research keywords
Highlights
AgentVidBench: a 100-question multi-hop video QA benchmark for MLLM agents
Judge-validation-first evaluation design
Recasting test-time prompt tuning as a calibration problem
The second 2026 arXiv paper reframes the test-time prompt tuning objective through the lens of calibration, extending the profile beyond evaluation into vision-language adaptation objectives; it awaits a full Evolve review.
Timeline
- 2026First catalogued release: AgentVidBench, a 100-question multi-hop video QA benchmark with LLM-judged trajectory scoring, appears on arXiv.
- 2026Posts a second arXiv paper, Rethinking the Test-Time Prompt Tuning Objective from the Perspective of Calibration, extending into vision-language adaptation objectives.
- 2026Both catalogued papers fall within computer vision / multimodal machine learning, making 2026 the only year captured by this profile so far.
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 dimensions are computed from a public bibliometric snapshot only; no ORCID or country is linked to this record.
- Only two papers have been reviewed; the test-time prompt tuning paper lacks a verdict, keywords, slug, and reproducibility assessment.
- Output, impact, and activity axes have no underlying data (citation counts, h-index, or multi-year volume), so they are left unscored rather than estimated.
- The reproducibility score of 100 rests on a single assessed paper (AgentVidBench) and may not generalize.