NLP researcher at Yonsei University: topic modeling and translation, now efficient inference and evaluation for LLMs
3
papers here
0
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
7
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 score is available yet. This axis would measure total publication volume, venue mix, and regularity; yearly counts look steady, but the venue-level and work-type data required for scoring is not in this record.
Impact—
No impact score is available yet. This axis would measure citation totals, field-normalized impact, and venue visibility; only h-index 7 and i10-index 4 are on record, which is too little to support a reliable score.
Focus—
No focus score is available yet. This axis would measure how concentrated output is across topics and venues; topic counts (NLP 67, topic modeling 63, translation 20, speech 19, dialogue 10) describe breadth but cannot yet be scored.
Collaboration51/100
Scoring 51 of 100, collaboration is moderate: the recent window records 12 distinct coauthors, with a tightly recurring core — Chanjun Park and Chanwoo Kim each appear twice in the frequent-coauthor list — alongside Doyun Kim, Yeo-Chan Yoon, Minsoo Song, Jungseob Lee, Seongtae Hong, Dongyub Jude Lee, Jaehyung Seo, and Heuiseok Lim. The pattern points to sustained output within one stable research group rather than broad cross-institutional networking.
Activity100/100
Activity scores a full 100 of 100, and the raw record backs it: 35 works over the last three years and yearly counts of 14 (2021), 16 (2022), 14 (2023), 14 (2024), 2 (2025), and 19 (2026). Apart from the thin 2025 figure, plausibly an indexing artifact, the cadence is steady, and 2026 is the strongest year on record.
Reproducibility—
No reproducibility score is available yet. This axis would measure code, data, and artifact release across works; only one reviewed paper carries a medium repro-readiness rating, which is far too little to characterize overall artifact-sharing practice.
Research summary
Sugyeong Eo is a researcher at Yonsei University, South Korea (ORCID 0000-0002-8008-6160), whose OpenAlex-resolved record is centered on natural language processing techniques (67 topic-assigned works) and topic modeling (63), with further activity in translation studies and practices (20), speech recognition and synthesis (19), and speech and dialogue systems (10). Output has been sustained: 14, 16, 14, and 14 works across 2021-2024, a thin 2 in 2025 (likely indexing lag), and 19 in 2026, with 35 works in the last three years. Citation indicators remain early-stage — h-index 7 and i10-index 4 — consistent with an ascending rather than mature citation record. Three 2026 arXiv preprints mark a pivot toward the LLM systems agenda: a one-pass block drafter for lossless speculative decoding in vision-language models (the one Evolve-reviewed paper, credited for strong engineering and statistics while flagging a capacity confound and the "entropy law" framing for scrutiny), red-teaming of autonomous skill generation in self-evolving agents, and a study of downward bias and role asymmetry in multi-agent LLM judges. All three list this researcher in a middle author position, and 12 distinct coauthors appear in the recent window.
Collaboration pattern
The collaboration footprint is compact and stable: 12 distinct coauthors in the recent window, with Chanjun Park and Chanwoo Kim each listed twice among frequent coauthors, alongside Doyun Kim, Yeo-Chan Yoon, Minsoo Song, Jungseob Lee, Seongtae Hong, Dongyub Jude Lee, Jaehyung Seo, and Heuiseok Lim. All three 2026 preprints are middle-author papers, consistent with contribution inside a persistent team rather than solo or corresponding-author-led projects.
Research keywords
Highlights
One-pass block drafting for lossless speculative decoding in vision-language models
Red-teaming autonomous skill generation in self-evolving agents
The 2026 preprint "EvoSkill Injection" examines how autonomous skill generation and evolution in self-evolving agents can be attacked, probing the safety surface of agent self-improvement. It is in the Evolve catalog but not yet reviewed.
Downward bias and role asymmetry in multi-agent LLM judges
The 2026 preprint "Beyond Consensus" studies how multi-agent LLM judges behave in subjective evaluation, documenting downward bias and role asymmetry. It is in the Evolve catalog but not yet reviewed.
Timeline
- 202114 works indexed — the first year of the provided counts window, starting a run of sustained output.
- 202216 works, the highest pre-2026 total in the record.
- 202314 works, keeping the yearly cadence steady.
- 202414 works — a fourth consecutive year in double digits.
- 2025Only 2 works indexed, an outlier most plausibly explained by indexing lag rather than a pause in output.
- 202619 works, the strongest year on record, including three arXiv preprints: lossless speculative decoding for VLMs, red-teaming of self-evolving agents, and multi-agent LLM judges.
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 indicators (h-index, i10-index, yearly counts, topics, coauthors) come from an OpenAlex-resolved identity; venue-level and citation-detail data are not in scope.
- The three 2026 catalog entries are arXiv preprints; only one has been reviewed by Evolve, and none lists this researcher as corresponding author.
- The 2025 count of 2 works likely reflects indexing lag rather than an actual output drop.
- Four radar axes (output, impact, focus, reproducibility) are unscored because underlying data is insufficient; no estimate is implied.