Systems researcher at UIUC working across LLM inference serving and LLM-based coding agents
2
papers here
2
reviewed by Evolve
0
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—
Insufficient data: no aggregate publication counts or venue history are available here. This axis would measure research output volume — papers per year and venue spread — across the full publication record, not just the two Evolve-reviewed 2026 papers.
Impact—
No citation counts, h-index, or downstream-adoption signals exist in the provided data, so this axis is intentionally unscored. It would measure influence — citations, reuse of the Orthrus scheduler or the STALE benchmark, and uptake by other LLM-serving and coding-agent teams.
Focus0/100
The focus score is 0 because the field distribution splits exactly evenly — 50% software engineering / LLM-based coding agents and 50% machine learning systems / LLM inference serving — across the two reviewed 2026 papers. Rather than one narrow niche, the record shows two parallel tracks that still converge on a single engineering question: making LLM-driven systems efficient and correct.
Collaboration41/100
The collaboration dimension scores 41 with 7 distinct coauthors recorded in this snapshot. Both 2026 papers list Park in the last author position, neither as corresponding author, suggesting a senior-contributor pattern. Recurring coauthor names include Haocheng Xia, Eugene Wu, Dohyun Park, Hubertus Franke, Daniel G. Waddington, and Swaminathan Sundararaman, though the data does not map individuals to specific papers.
Activity—
The snapshot provides no publication-timestamp series, so recent-year cadence cannot be scored. This axis would measure research activity — number of outputs per recent year, submission rhythm, and continuity — requiring a longer publication history than the two 2026 arXiv entries reviewed here.
Reproducibility78/100
Reproducibility is the strongest measured axis: a mean readiness score of 78 across the 2 papers assessed. The Orthrus serving paper is rated high readiness, consistent with systems work built around controlled comparisons, while the STALE benchmark paper is rated medium. Evolve's review praises the benchmark's honesty and rigor — mined PR pairs almost never interfere after the grading fix — indicating careful evaluation practice.
Research summary
Yongjoo Park is a researcher at the University of Illinois Urbana-Champaign whose catalogued work sits at the junction of two systems-oriented fields: machine learning systems for LLM inference serving, and software engineering for LLM-based coding agents. The profile is name-resolved from public bibliographic data tied to ORCID 0000-0003-3786-6214 and covers two 2026 arXiv papers, on both of which Park appears in the last author position, neither flagged as corresponding author. The first, 'Efficient Iterative Retrieval with Heterogeneous Batching', presents Orthrus, which unifies embedding and generation batching through chunked pooling and credit scheduling; reported gains run from 4.5x in controlled settings to roughly 14% over the strongest RAG baseline, and its reproducibility readiness is rated high. The second, 'Passes Alone, Fails Together', introduces the STALE benchmark for semantic coordination in parallel LLM-agent development; after a grading fix, mined pull-request pairs almost never interfere, and scripted Django edits fail 97% of the time — a result Evolve's review calls honest and rigorous, though narrow and idealized. Field distribution splits evenly (50/50) across the two areas, yielding a focus score of 0, while collaboration scores 41 over 7 distinct coauthors and reproducibility averages 78 across the 2 assessed papers. Output, impact, and activity axes lack underlying data and remain intentionally unscored.
Collaboration pattern
The snapshot records 7 distinct coauthors, yielding a collaboration score of 41. Recurring names include Haocheng Xia, Eugene Wu, Dohyun Park, Hubertus Franke, Daniel G. Waddington, and Swaminathan Sundararaman, though the data does not map them to individual papers. Park sits in the last author position on both 2026 papers without corresponding-author flagging — a senior-contributor pattern across both output tracks.
Research keywords
Highlights
Orthrus: unified batching for embedding and generation in iterative RAG
STALE benchmark finds parallel LLM agents rarely collide semantically
Senior-author positioning across both 2026 releases
On both catalogued papers Park appears in the last author position without corresponding-author flagging, and the pair spans two fields split 50/50, with reproducibility readiness rated high and medium respectively within the 78-point mean.
Timeline
- 2026arXiv release of 'Efficient Iterative Retrieval with Heterogeneous Batching', presenting the Orthrus approach that unifies embedding and generation batching via chunked pooling and credit scheduling.
- 2026arXiv release of 'Passes Alone, Fails Together', the STALE benchmark for semantic coordination in parallel LLM-agent development.
- 2026Both catalogued outputs land in the same year with last-author, non-corresponding positioning, spanning LLM inference serving and LLM-based coding agents.
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.
- Profile resolved from public bibliographic records via ORCID 0000-0003-3786-6214; 'Yongjoo Park' is not a rare name, so identity is matched bibliographically rather than biographically.
- Dimension scores come from a small bibliometric snapshot (7 distinct coauthors, 2 reviewed papers); the output, impact, and activity axes have no underlying data and are intentionally unscored.
- Coverage is limited to the two Evolve-reviewed 2026 arXiv entries; any other publications, citations, or venue history fall outside this snapshot.
- The frequent-coauthor list includes the researcher's own name and was excluded from collaboration counts; the country field is missing in the source data.