NLP researcher at Sungkyunkwan University — efficient LLM reasoning, judge robustness, computational argumentation
6
papers here
4
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
16
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—
Not yet scored: per-work publication volume is incomplete in the source record, so no reliable output estimate exists. This axis would measure annual and cumulative paper counts — how much the researcher publishes over time.
Impact—
Citation-based influence is not scored: citation linkage for this record is insufficient despite a listed h-index of 16. This axis would measure citation counts, citation trajectory, and uptake of the researcher's work by others.
Focus0/100
Scoring 0, topical focus is highly dispersed. The five largest OpenAlex topic clusters — veterinary orthopedics (48), shoulder injury (44), topic modeling (38), seismic performance (34), and customer service quality (31) — span unrelated domains, while the recent catalog concentrates on three NLP fields at one third each. The long-tail spread likely signals record-level name ambiguity.
Collaboration78/100
Scoring 53, the collaboration profile shows 13 distinct coauthors in the last three years plus a recurring twelve-name coauthor list (including DongHyun Ryu, Yohan Jo, and Joonsuk Park). Across the three reviewed 2026 papers Bak is consistently last author though not corresponding, pointing to a senior supervisory position within small NLP teams rather than a large distributed network.
Activity100/100
Scoring a full 100, recent activity is exceptional: 44 works in the last three years, matching the annual counts of 16 (2024), 11 (2025), and 17 (2026). This steady arXiv cadence — including three reviewed preprints in 2026 alone — shows sustained, high-frequency current productivity.
Reproducibility89/100
Scoring 85 from three papers with reproducibility assessments (reproMean 85, reproPapers 3), this is a clear strength. When2Think and the LLM-as-a-judge noise study are rated high readiness, while PTCG is rated medium. Evolve verdicts praise the controlled methodology and flag thin statistics or ablation gaps rather than reproducibility barriers.
Research summary
JinYeong Bak is a researcher at Sungkyunkwan University (South Korea) whose recent catalog output sits in natural language processing, machine learning for efficient LLM reasoning, and computational argumentation — the three fields are evenly split in the profile's field distribution (one third each). The Evolve catalog holds four 2026 arXiv papers; on all three reviewed ones Bak appears as last author (non-corresponding), consistent with a senior, supervision-oriented role. The work spans difficulty-aware length control for hybrid reasoning (When2Think), a controlled study showing that noisy judged text inflates LLM-as-a-judge bias measurements — with fabrication-over-erasure ratios up to 120× — and persona-guided tree-based counterargument generation (PTCG). Bibliometric aggregates (OpenAlex) record an h-index of 16, an i10-index of 26, 44 works over the last three years, and 13 distinct recent coauthors; recent annual counts (16 in 2024, 11 in 2025, 17 in 2026) support a high activity rating. Reproducibility is a strength: mean readiness of 85 across three reviewed papers, two rated high. The aggregated record, however, spans 1984–2026 and includes topics far outside NLP, so long-run counts likely reflect name ambiguity and should be read cautiously.
Collaboration pattern
Collaboration is compact and stable: 13 distinct coauthors in the last three years and a recurring group including DongHyun Ryu, Jaehyeok Lee, YeongJun Hwang, Jaejun Shim, HyunJin Kim, Yohan Jo, and Joonsuk Park. On all three reviewed 2026 papers Bak is the last author without corresponding status, a pattern consistent with senior supervision of student-led arXiv projects rather than large multi-institution consortia. The collaboration dimension scores 53.
Research keywords
Highlights
Exposing fragility in LLM-as-a-judge bias measurement
Difficulty-aware length control for efficient hybrid reasoning
Persona-guided tree search for counterargument generation
Timeline
- 2026Published PTCG, combining persona selection with Tree-of-Thoughts pruning for counterargument generation (arXiv, last author).
- 2026Published the controlled study on LLM-as-a-judge bias measurement under noisy text, documenting fabrication-over-erasure ratios up to 120×.
- 2026Published When2Think on difficulty-aware length control for efficient hybrid reasoning models.
- 2026Appeared on CHARM, a character-hallucination benchmark for multicultural role play (not yet reviewed by Evolve).
Papers in the Evolve catalog
When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models
When Noise Fabricates Bias: The Fragility of LLM-as-a-Judge Bias Measurement under Noisy Text
PTCG: Persona-guided Tree-based Counterargument Generation
CHARM: Character Hallucination for Multicultural Role Play Benchmark
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 aggregates come from OpenAlex only; the output and impact dimensions are unscored due to insufficient underlying data.
- The aggregated record spans 1984–2026 and includes topics far outside NLP (e.g., veterinary orthopedics, seismic analysis), so long-run counts and topic spread likely reflect name ambiguity and should be read cautiously.
- Only three of the four catalog papers carry Evolve review verdicts; CHARM has no verdict, keywords, or slug yet.
- The reproducibility score (85) is based on only three assessed papers.