Dong Ho Kang

Research profile

Research summary

Dong Ho Kang is a researcher at Kookmin University working in machine learning evaluation methodology, with a concentration on large language model (LLM) evaluation. The profile is built from two 2026 arXiv preprints in the Evolve catalog, on both of which Kang holds the last author position. The first, 'Auditing Proxy-Based Validation Across Text Spans', audits proxy-based validation across text spans and shows that same-span proxy agreement can amount to mere surface containment — a construct-validity concern relevant to LLM-as-a-judge, model routing, and refusal detection; Evolve credits the careful method while noting 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 functions as an audit diagnostic rather than a repair, documenting format effects and slice gains alongside their limits. The field distribution is an even 0.5/0.5 split between LLM evaluation methodology and ML evaluation methodology/NLP. Both papers are rated high for reproducibility readiness, with a reproducibility mean of 100 across 2 scored papers. Collaboration is compact — 3 distinct coauthors, with Daein Weon recurring as the key partner. Output, impact, and activity metrics are not yet available and are intentionally not scored.

Research keywords

Highlights

• Proxy Agreement Across Text Spans Can Be Mere Surface Containment

  Last-authored 2026 arXiv study auditing proxy-based validation across text spans for LLM evaluation. The span-boundary audit finds that same-span proxy agreement can reduce to surface containment, with implications for LLM-as-a-judge, model routing, and refusal detection; Evolve rates the method solid but notes evidence is limited to diagnostic scores on one model pair.

• Residualization of LLM Eval Scores Is an Audit Diagnostic, Not a Repair

  The second 2026 paper argues that post-hoc residualization of LLM evaluation scores should be treated as an audit diagnostic rather than a fix, documenting format effects and slice gains together with their limits; Evolve praises the careful, honest scoping while noting the central harm is shown only synthetically.

• High Reproducibility Readiness Across Both Papers

  Both catalogued papers carry 'high' reproducibility readiness, yielding a reproducibility mean of 100 across 2 scored papers — the strongest quantitative signal in this profile, reflecting carefully scoped methods and explicit self-stated evidence limits.

Timeline

1. 2026First catalogued arXiv preprint appears: 'Auditing Proxy-Based Validation Across Text Spans', with Dong Ho Kang in the last-author position.
2. 2026Second arXiv preprint, 'When Residualization Helps an Audit: Format Effects, Slice Gains, and Their Limits', argues residualization is an audit diagnostic rather than a repair.
3. 2026Both 2026 papers receive 'high' reproducibility readiness ratings (reproMean 100 across 2 papers), anchoring the year's output in LLM evaluation methodology.

Papers in the Evolve catalog

• 2026-09-22

  Auditing Proxy-Based Validation Across Text Spans

• 2026-09-21

  When Residualization Helps an Audit: Format Effects, Slice Gains, and Their Limits

Frequent co-authors