Geewook Kim
Korea Advanced Institute of Science and Technology · South Korea
KAIST researcher probing multimodal LLMs: drone-control agents, reproducible evaluation, attention-sink diagnosis.
3
papers here
0
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Output is unscored: total publication count, venue mix, and the full yearly distribution are not reliably available in this OpenAlex record. Once collected, this axis would measure productivity volume across the researcher's whole career.
Impact—
Impact is unscored: citation-based measures are unreliable here — the snapshot reports h-index 0 and i10-index 0, which most likely reflects incomplete citation linkage rather than zero influence. This axis would measure scholarly uptake and citation-based reach.
Focus—
Focus is unscored: topic labels exist — multimodal machine learning applications (3), topic modeling (2), and three single-count topics — but the sample is too thin to compute topical concentration. This axis would measure how tightly work clusters around core themes.
Collaboration61/100
Collaboration scores 61/100. The OpenAlex snapshot lists 20 distinct coauthors overall, and 12 of them recur in the Evolve-reviewed papers — Dohoon Ryu, Minsik Choi, Young Geun Kim, Hodong Lee, Sanghee Park and others. Last-author placement on two of the three 2026 preprints, with no corresponding-author role, fits a senior collaborator anchoring a rotating team rather than a solo writer.
Activity40/100
Activity scores 40/100. The snapshot records 3 works over the last three years, all dated 2026 — countsByYear contains a single 2026 entry with 3 outputs — and all three are arXiv preprints in the Evolve catalog. The pattern suggests a compact, current burst of evaluation-focused output rather than a sustained multi-year publication stream.
Reproducibility—
Reproducibility is unscored: systematic evidence on code release, artifact availability, and independent replication has not been collected, despite Evolve's per-paper readiness ratings (one high, two medium). This axis would measure how readily reported results can be reproduced.
Research summary
Geewook Kim is a KAIST-affiliated researcher whose recent work concentrates on evaluating multimodal foundation models. The Evolve catalog lists three 2026 arXiv preprints: a benchmark that flies nine multimodal LLMs as zero-shot vision-language-action agents for drone commanding, approaching, tracking and searching; a composable evaluation system for reproducible omni-modal foundation-model evaluation, anchored by a CPU-side answer verifier and LLM-as-a-judge; and an attention-sink diagnosis showing that a cheap base-LLM scalar, Sink Strength, predicts post-adaptation text capability loss at Spearman ρ=0.97 (n=6). Kim holds the last author position on the first two and a middle position on the third, never corresponding author — a placement consistent with a senior-contributor or supervisory role. OpenAlex records 20 distinct coauthors and 3 works in the last three years, all dated 2026, with topic clusters in multimodal machine learning applications (3), topic modeling (2), robot manipulation and learning, domain adaptation and few-shot learning, and scientific computing. Radar scores collaboration at 61/100 and activity at 40/100; output, impact, focus and reproducibility remain unscored because the underlying data is insufficient.
Collaboration pattern
The collaboration record centers on a stable KAIST-affiliated cluster: OpenAlex lists 20 distinct coauthors, and 12 recur across the reviewed papers, including Dohoon Ryu, Minsik Choi, Young Geun Kim, Hodong Lee and Sanghee Park. Kim consistently takes a late author position — last on two 2026 preprints, middle on one, never corresponding — consistent with a senior contributor or advisor rather than a hands-on first author.
Research keywords
Highlights
Nine MLLMs tested as zero-shot drone pilots
A composable system for reproducible omni-modal evaluation
Sink Strength predicts vision-language text loss
Timeline
- 2026Last-authored a benchmark flying nine multimodal LLMs as zero-shot drone-control agents in simulation.
- 2026Last-authored the composable evaluation system for reproducible omni-modal foundation-model evaluation.
- 2026Co-authored the attention-sink diagnosis linking Sink Strength to post-adaptation text capability loss (ρ=0.97, n=6).
Papers in the Evolve catalog
Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching
A Composable Evaluation System for Reproducible Omni-Modal Foundation Model Evaluation
Text Capability Loss in Vision-Language Adaptation: An Attention-Sink Diagnosis
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.
- Bibliometrics come from an OpenAlex snapshot (identity: openalex) only; h-index and i10-index are reported as 0, indicating thin citation coverage rather than a measured absence of impact.
- No ORCID is linked; the profile is name-resolved from public bibliographic data, so disambiguation against other researchers with the same name is not fully guaranteed.
- Evidence covers three 2026 arXiv preprints reviewed by Evolve; earlier works, peer-reviewed venues, and citation records are outside the current scope.
- Four radar axes (output, impact, focus, reproducibility) are unscored due to insufficient data; they are reported as missing, not as zeros.