Hyunwoo J. Kim

Korea Advanced Institute of Science and Technology · United States

Unclaimed · AI-structured profileOpenAlex

LLM researcher spanning multi-agent prompt optimization, agent-based social simulation, and pragmatic LLM benchmarking

papers here

reviewed by Evolve

first-author

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—

No output score is available: the underlying record lacks sufficient publication-volume data to score this axis. Once fuller venue and preprint counts are ingested, this dimension would measure total research output and how it grows over time.

Impact—

No impact score is available: citation-based measures are missing or uninformative in the current record, so the axis is left unscored. With better coverage it would capture citation influence, h-index-style accumulation, and uptake of the cataloged papers.

Focus0/100

The focus score is 0/100, and the field distribution backs that up: the three cataloged 2026 papers split evenly (0.333 each) across LLM-based multi-agent systems, computational social science, and LLM evaluation. Topic counts echo the breadth — Topic Modeling, Natural Language Processing Techniques, and Multi-Agent Systems and Negotiation appear twice each, with Ethics and Social Impacts of AI and Speech and Dialogue Systems once each. This is a deliberately broad LLM-systems portfolio rather than a single-track specialty.

Collaboration64/100

Collaboration scores 64/100. The OpenAlex record shows 24 distinct coauthors in the recent window, and the Evolve catalog lists 12 frequent collaborators — Chaemin Jang, Junsik Min, Jaewoo Choi, Donggyu Lee, Haiin Lee, Junyoung Park, Namhee Kim, Jungwon Kim, Juho Kim, Nuri Kim, Jihee Kim, and Akhila Yerukola. Across the three papers this researcher's author position varies (last on AgentGrad, middle on the others), consistent with a rotating, team-based arXiv workflow.

Activity40/100

Activity scores 40/100. The record's countsByYear shows 3 works in 2026, and recent3yWorks is also 3 — all currently visible output lands in a single year and consists of arXiv preprints. The moderate score therefore reflects a short observation window more than it reflects the actual research tempo.

Reproducibility70/100

Reproducibility scores 70/100, the mean across 3 papers carrying Evolve reproducibility assessments. Per paper, readiness is medium for AgentGrad and XYBench and high for the simulated-deliberation study, which Evolve credits for strong controls despite a descriptive-statistics-only analysis. No cataloged paper is rated low, suggesting experimental setups and protocols that support replication.

Research summary

Hyunwoo J. Kim is a researcher affiliated (per the underlying record) with the Korea Advanced Institute of Science and Technology, working across three evenly weighted strands of large language model research: prompt optimization for multi-agent systems, agent-based social simulation, and LLM evaluation. The Evolve catalog holds three 2026 arXiv papers. AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems (last author) applies sequential intervention and gradient clustering to textual-gradient prompt optimization, with strong reported gains and a 2.5x speedup, though Evolve notes its SOTA claim conflicts with the paper's own Table 2. From Simulated Citizens to Simulated Deliberation (middle author) shows that LLM persona deliberation fails population representation, with a 29 percentage-point gap, and that stances shift without genuine peer exchange. XYBench (middle author) is an 8,115-query benchmark on pragmatic responses to queries carrying misconceptions, finding that models answer literal requests but fail pragmatic redirection. The record lists 24 distinct coauthors and 12 frequent collaborators, with the three fields splitting one third each; collaboration scores 64/100, reproducibility 70/100, activity 40/100, and focus 0/100, while output and impact lack sufficient data to score. All three outputs are arXiv preprints with no corresponding-author role marked, so this reads as an active, deliberately broad LLM-systems program whose bibliometric footprint has not yet accumulated.

Collaboration pattern

Collaboration scores 64/100, with 24 distinct coauthors in the recent OpenAlex window and 12 recurring collaborators — Chaemin Jang, Junsik Min, Jaewoo Choi, Donggyu Lee, Haiin Lee, Junyoung Park, Namhee Kim, Jungwon Kim, Juho Kim, Nuri Kim, Jihee Kim, and Akhila Yerukola. The three cataloged papers all carry team-sized author lists, and author position varies (last on AgentGrad, middle elsewhere), suggesting a flexible, project-based collaboration pattern.

Research keywords

prompt optimizationmulti-agent systemstextual gradientsfailure attributioncredit assignmentLLM agent simulationmulti-agent deliberationpersona agentsopinion dynamicsLLM evaluationpragmaticsmisconceptions

Highlights

AgentGrad: intervention-guided prompt optimization for multi-agent systems
As last author on this 2026 arXiv paper, contributes a method that treats multi-agent failure attribution as sequential intervention and clusters textual gradients to optimize agent prompts; Evolve reports strong gains and a 2.5x speedup while flagging that the SOTA claim conflicts with the paper's Table 2. Reproducibility readiness: medium.
Evidence that LLM persona deliberation breaks population representation
Middle author on this computational social science study, which finds that simulated citizen personas deliberating in groups leave a 29 percentage-point representation gap and shift stances without genuine peer exchange; Evolve credits the strong controls and survey benchmarking while noting the analysis is descriptive. Reproducibility readiness: high.
XYBench: LLMs answer literal requests but fail pragmatic redirection
Middle author on this NLP evaluation paper, which introduces an 8,115-query benchmark for cooperative responses to queries containing misconceptions; models largely satisfy the literal ask yet miss the pragmatic redirect. Evolve judges the benchmark well validated but cautions that LLM-built ground truth tempers the gap claims. Reproducibility readiness: medium.

Timeline

2026Contributes to XYBench, an 8,115-query benchmark testing whether LLMs respond pragmatically to queries carrying misconceptions (arXiv preprint, middle author).
2026Co-authors a computational social science study diagnosing representation and interaction failures in LLM persona deliberation (arXiv preprint, middle author).
2026Introduces AgentGrad, an intervention-guided prompt-optimization method for multi-agent systems (arXiv preprint, last author).

Papers in the Evolve catalog

Frequent co-authors

Chaemin Jang×1 Junsik Min×1 Jaewoo Choi×1 Donggyu Lee×1 Haiin Lee×1 Junyoung Park×1 Namhee Kim×1 Jungwon Kim×1 Juho Kim×1 Nuri Kim×1 Jihee Kim×1 Akhila Yerukola×1

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.

Bibliometrics are drawn from OpenAlex only, and citation coverage for this record is thin — h-index and i10-index are listed as 0, which we treat as insufficient citation data rather than a measured ranking.
The output and impact axes are unscored (dimMissing) because the underlying publication and citation data are insufficient.
"Hyunwoo J. Kim" is a common Korean name; this profile is name-resolved from public bibliographic data (OpenAlex identity) plus the three cataloged 2026 arXiv papers, so verify identity before citing.
Institution and country fields are reproduced as listed in the record (KAIST; country "US"), which appear inconsistent and are unverified.