Kookmin University researcher auditing proxy validity and residualization in LLM evaluation methodology
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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 publication-volume metrics are available in the current snapshot, so output is unscored. The axis would measure total production — paper counts, venue distribution, and the share of first-authored versus co-authored work — over the researcher's full publication record.
Impact—
Citation evidence is not yet available, so impact is unscored. This axis would measure citations to the reviewed papers, their growth over time, and downstream uptake of the proposed audit methods by other LLM-evaluation studies.
Focus0/100
Focus scores 0/100 because the field distribution is evenly split: 50% machine learning (LLM evaluation methodology) and 50% machine learning evaluation methodology / NLP, with no dominant field. Substantively the halves are adjacent — both concern how LLM evaluation proxies and judge scores are validated — so the split likely reflects labeling granularity more than genuine topic dispersion, and the record still spans proxy audits, residualization, routing, and refusal detection.
Collaboration28/100
The bibliometric record shows a compact collaboration footprint: only 3 distinct co-authors appear across the reviewed papers, for a collaboration score of 28/100. The dominant repeat partner is Dong Ho Kang, who appears twice in the frequent-coauthor list. Both catalog papers are first-authored and not corresponding, consistent with a small-team style in which Weon leads the analysis and writing rather than joining large consortia.
Activity—
Longitudinal data is insufficient: the visible record covers only 2026 papers, so activity is unscored. The axis would measure publication cadence and continuity across recent years, such as papers per year and gaps between releases.
Reproducibility100/100
Reproducibility is the strongest measured axis: the two reviewed papers average 100 (reproMean 100 across 2 reproPapers), and both carry high reproducibility readiness in the Evolve catalog. The audits rest on transparent span-level diagnostics — for example, diagnostic scores on a single model pair — which the reviews describe as carefully scoped and honest, consistent with deliberately reproducible evaluation-methodology work, albeit on a narrow evidence base.
Research summary
Daein Weon is a researcher at Kookmin University working in machine learning evaluation methodology for large language models. This is an early-stage profile built from two 2026 arXiv preprints, both first-authored and not corresponding. The first, "Auditing Proxy-Based Validation Across Text Spans", asks whether agreement between a cheap proxy and a reference signal over text spans reflects genuine construct validity or mere surface containment; the review praises the careful span-boundary audit while noting the evidence is limited to diagnostic scores on one model pair. The second, "When Residualization Helps an Audit", argues that post-hoc residualization of LLM evaluation scores should be treated as an audit diagnostic rather than a repair, examining format effects and slice gains with, per the review, honest but partly synthetic evidence. Together they define a coherent agenda: stress-testing the validity of proxies, judges, and score adjustments used in LLM evaluation, spanning LLM-as-a-judge pipelines, reward models, model routing, refusal detection, and annotation artifacts. The measured bibliometric picture is narrow but distinctive: the field splits evenly between LLM evaluation methodology and evaluation methodology/NLP (0.5/0.5), the collaboration score is 28 on the strength of only 3 distinct co-authors anchored by repeat collaborator Dong Ho Kang, and reproducibility reaches a mean of 100 across 2 papers, both rated high readiness. Output, impact, and activity remain unscored for lack of data.
Collaboration pattern
The visible collaboration pattern is a tight small-team core: 3 distinct co-authors in total, with Dong Ho Kang the only repeated partner, appearing twice in the frequent-coauthor list. Both catalog papers are first-authored by Weon and neither is marked corresponding, consistent with Weon leading the analysis and writing. No large-consortium or multi-institution collaborations are visible in the current record.
Research keywords
Highlights
Showing that same-span proxy agreement can be mere surface containment
Reframing residualization of LLM eval scores as an audit diagnostic, not a repair
Timeline
- 2026Published the first-authored arXiv preprint "Auditing Proxy-Based Validation Across Text Spans", auditing whether proxy agreement over text spans reflects construct validity or surface containment.
- 2026Published "When Residualization Helps an Audit", arguing that post-hoc residualization of LLM evaluation scores is an audit diagnostic rather than a repair.
- 2026Both 2026 papers entered the Evolve catalog with high reproducibility readiness, yielding a reproducibility mean of 100 across 2 reviewed papers.
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.
- Radar scores and co-author counts come from public bibliometric data and cover only the 2 papers Evolve has reviewed; no full publication record is included in this profile.
- The output, impact, and activity dimensions are unscored because citation counts, publication volume, and multi-year cadence data are not available in this snapshot.
- No ORCID or country data was provided; this profile is name-resolved from public bibliographic data, and the name Daein Weon may be shared by more than one person.
- The frequent-coauthor list contains the researcher's own name and a duplicate entry for Dong Ho Kang, indicating entity-resolution noise in the underlying data.