First-author researcher in LLM alignment safety and model compression; rigorous, measurement-driven evaluation studies.
2
papers here
2
reviewed by Evolve
2
first-author
1
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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: the snapshot contains no publication-count or output-volume metrics beyond the two reviewed papers, so this axis cannot be computed. It would measure total and recent scholarly output volume.
Impact—
Not yet scored: no citation counts, h-index, or downstream-adoption signals appear in the input data. This axis would measure how widely the work is cited and built upon by other researchers.
Focus0/100
The focus axis currently scores 0, reflecting a field distribution split exactly 50/50 between information retrieval / NLP model compression and natural language processing / LLM alignment safety. With only two reviewed papers, one in each area, the profile does not yet show a single dominant specialization; both strands share a measurement-driven evaluation style, and a clearer focus should emerge as more output is indexed.
Collaboration0/100
The collaboration axis scores 0 because the snapshot records zero distinct co-authors and an empty frequent-coauthor list. Both reviewed papers list Han as first author — one also as corresponding author — which is consistent with a small-team or solo working mode, but with no co-author metadata captured, the actual collaboration pattern cannot yet be characterized from this data.
Activity—
Not yet scored: the input provides no publication-frequency or per-year output series, so research tempo cannot be computed. This axis would measure publication cadence and sustained research activity over time.
Reproducibility78/100
The reproducibility axis scores 78, the mean over the two reviewed papers. The post-training quantization study carries high reproducibility readiness and is described by Evolve as unusually transparent, while the alignment study is rated medium, reflecting its single-model, three-seed design. Together they point to an unusually careful, measurement-first evaluation practice.
Research summary
Hyojung Han is an early-stage researcher working at the intersection of NLP model reliability and alignment safety. The current evidence base is two 2026 arXiv papers in the Evolve catalog, both with Han as first author, one also as corresponding author. The profile's field distribution splits evenly — 0.5 to information retrieval / NLP model compression and 0.5 to natural language processing / LLM alignment safety — so no single specialization dominates yet. The first paper, 'Where Post-Training Quantization Breaks Text Embedders,' maps where standard quantization heuristics fail to transfer across four embedder families; Evolve characterizes it as a rigorous, unusually transparent negative-result study with narrow scope and rates its reproducibility readiness high. The second, 'Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model,' reports rigorously controlled evidence that style-only post-training shifts bias exposure and disclosure through answer propensity and verbosity, with the caveat of a single model and three seeds. The reproducibility mean across these two papers is 78. No institutional affiliation, country, ORCID, or co-author network is captured in the snapshot, and citation-based output and impact metrics are absent, so this profile should be read as an early-stage, name-resolved sketch built strictly from reviewed work.
Collaboration pattern
No co-author network is captured in the current snapshot: the distinct co-author count is zero and the frequent-coauthor list is empty, so no key collaborators can be named. Both reviewed papers are first-authored, with the Korean alignment study also listing Han as corresponding author, which is consistent with leading small, evaluation-focused projects. Collaboration patterns should be re-assessed once author metadata and a larger paper sample become available.
Research keywords
Highlights
Mapping where post-training quantization breaks text embedders
Linking response-style alignment to off-target bias and disclosure shifts
A reproducibility-first evaluation style across two fields
Across the two reviewed papers the reproducibility mean is 78 (one high, one medium readiness), and the field distribution splits evenly between LLM alignment safety and NLP model compression — a consistent, measurement-driven style of negative results and controlled experiments.
Timeline
- 2026Published first-author arXiv study 'Where Post-Training Quantization Breaks Text Embedders,' mapping quantization failures across four embedder families; Evolve rates its reproducibility readiness high.
- 2026Published first- and corresponding-author arXiv study 'Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model,' tying style-only post-training to shifted bias exposure and disclosure.
- 2026Both papers entered the Evolve review catalog, establishing a two-paper, two-field early profile with a reproducibility mean of 78 and an even field split.
Papers in the Evolve catalog
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 scores are computed from a limited OpenAlex-based snapshot; only two reviewed papers underpin this profile.
- No institutional affiliation, country, or ORCID is available; the profile is name-resolved from public bibliographic data and 'Hyojung Han' may match multiple researchers.
- Output, impact, and activity axes lack underlying data (no citation counts or publication-volume series) and are left unscored rather than estimated.
- The zero distinct-coauthor count reflects missing author metadata in the snapshot, not a confirmed solo working style.