Early-career NLP researcher red-teaming self-evolving agents, auditing multi-agent LLM judges and MCQA benchmarks
3
papers here
0
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Insufficient data: the snapshot contains no cumulative publication count beyond three 2026 arXiv preprints (recent3yWorks = 3). This axis would measure total and per-year scholarly output; it is left unscored rather than estimated.
Impact—
Insufficient data: OpenAlex reports h-index 0 and i10-index 0 with no citation counts. This axis would measure citation-based influence and uptake of the work; it is left unscored rather than estimated.
Focus—
Insufficient data: OpenAlex lists five topic labels (topic modeling, NLP techniques, adversarial robustness among them) but coverage is too sparse to compute concentration. This axis would measure how focused the publication portfolio is.
Collaboration39/100
The collaboration score is 39, computed from six distinct coauthors in the snapshot. The three 2026 preprints share a recurring core team — frequent collaborators include Sugyeong Eo, Minsoo Song, Doyun Kim, Chanwoo Kim, and Yeo-Chan Yoon — and Park sits in the last author position on all three, a position conventionally associated with senior roles, though none is flagged as corresponding author. The small, stable group suggests tight lab-style collaboration rather than broad external networks.
Activity40/100
The activity score is 40. The snapshot's countsByYear contains a single entry — 3 works in 2026 — and recent3yWorks is also 3, meaning all indexed output is concentrated in one very recent year, all as arXiv preprints. No earlier years appear in the record, so this reflects a recent burst of submissions rather than a sustained multi-year cadence; the score should firm up as coverage extends.
Reproducibility—
Insufficient data: no code, dataset, or artifact links are recorded for the three preprints, and no Evolve reproducibility verdicts exist yet. This axis would measure whether papers release reusable code and data; it is left unscored.
Research summary
Chanjun Park is an early-stage researcher working at the intersection of natural language processing, adversarial machine learning, and multi-agent LLM systems. The Evolve catalog lists three 2026 arXiv preprints, each with Park in the last author position: EvoSkill Injection, which red-teams autonomous skill generation and evolution in self-evolving agents; Beyond Consensus, which documents downward bias and role asymmetry in multi-agent LLM judges for subjective evaluation; and Auditing MCQA Benchmarks through Probability Landscapes, which probes what multiple-choice benchmark scores conceal. OpenAlex topic assignments reinforce this evaluation-and-safety orientation: Topic Modeling and Natural Language Processing Techniques (2 works each), Adversarial Robustness in Machine Learning (2), plus Multi-Agent Systems and Negotiation and Information and Cyber Security (1 each). Coverage is thin: the snapshot reports h-index 0 and i10-index 0, three works in the recent three-year window, and six distinct coauthors, so the output, impact, focus, and reproducibility axes are marked missing rather than scored, while collaboration and activity stand at 39 and 40. The through-line is a coherent trustworthy-LLM agenda — attacking agent skill evolution, stress-testing judge panels, and auditing benchmarks — whose citation footprint has not yet accumulated in indexed sources; read this profile as an early-stage snapshot rather than a mature record.
Collaboration pattern
Park works with a small, stable core team: the snapshot records six distinct coauthors, and frequent collaborators include Sugyeong Eo, Minsoo Song, Doyun Kim, Chanwoo Kim, and Yeo-Chan Yoon. All three 2026 arXiv preprints share this group, with Park in the last author position on each and no corresponding-author flag recorded. The pattern points to a tight lab-style cluster rather than broad external networks.
Research keywords
Highlights
Red-Teaming Self-Evolving Agents
As last author of 'EvoSkill Injection' (arXiv, 2026), Park's team attacks how autonomous skill generation and evolution in self-evolving agents can be exploited, probing a core safety surface of agentic LLMs.
Exposing Multi-Agent Judge Bias
'Beyond Consensus' (2026) documents downward bias and role asymmetry in multi-agent LLM judges used for subjective evaluation, cautioning against treating panel consensus as fairness or accuracy.
Auditing MCQA Benchmarks
In 'Auditing MCQA Benchmarks through Probability Landscapes' (2026), the team examines models' full probability landscapes to audit what multiple-choice benchmark scores hide beneath headline accuracy.
Timeline
- 2026Publishes 'EvoSkill Injection' on arXiv, red-teaming autonomous skill generation and evolution in self-evolving agents, in the last author position.
- 2026Publishes 'Beyond Consensus' on arXiv, analyzing downward bias and role asymmetry in multi-agent LLM judges for subjective evaluation.
- 2026Publishes 'Auditing MCQA Benchmarks through Probability Landscapes' on arXiv, completing a three-paper 2026 run on trustworthy LLM evaluation.
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.
- Bibliometric scores, topic counts, and identity resolution come from an OpenAlex snapshot; h-index and i10-index are reported there as 0.
- Institution, country, and ORCID are unavailable; this profile is name-resolved from public bibliographic data and 'Chanjun Park' may be shared by multiple researchers.
- Only three 2026 arXiv preprints are in scope; no venue, peer-review status, or Evolve review verdicts are recorded yet.
- Reproducibility signals (code and dataset links) are absent from the data, so artifact availability cannot be assessed.