Qualcomm researcher working across efficient LLM inference and Korean visual speech recognition
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) 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 computed yet: publication and preprint volume data for this profile is insufficient. This axis would measure total papers over time, first/corresponding-author share, and yearly productivity once the bibliographic record is complete enough to score.
Impact—
Impact cannot be scored because citation data is not available in the current snapshot. This axis would aggregate citations, citations per paper, and uptake of the KV-cache eviction and Korean lipreading methods by other work once citation records accumulate.
Focus0/100
Evolve's focus axis scores 0/100, driven by the raw field distribution: 50% machine learning (efficient LLM inference) and 50% speech processing (audio-visual speech recognition). Across the two catalogued 2026 papers, the work splits evenly between KV-cache eviction for LLM inference and Korean lipreading on OLKAVS, so the profile reads as two parallel research lines rather than one concentrated topic.
Collaboration41/100
The collaboration axis scores 41/100, based on 7 distinct co-authors in the underlying data. The recurring pool is small and consistent with an industry lab: Jungwook Choi, Mingu Lee, Se Un Park, Hakjun Kim and Taehoon Roh. Both 2026 papers are arXiv preprints from this tight group — one first-authored, one last-authored — indicating stable internal partnerships rather than broad external networking.
Activity—
Activity is unscored because a recent-publication cadence cannot be established from the current record, which anchors on two 2026 arXiv preprints. This axis would measure publication frequency and continuity across recent years once a fuller publication history is linked.
Reproducibility55/100
The reproducibility axis scores 55/100, a mean computed over 2 papers, each rated medium for repro readiness. Evolve's review notes support this middle band: the Korean lipreading paper runs rigorous paired evaluation but lacks prior-method baselines, while ValueDiff reports strong gains on sink-suppressed models without significance testing on its main results — methods are described but verification gaps remain.
Research summary
Junyoung Park is a researcher at Qualcomm (United States) whose public bibliographic footprint in the Evolve catalog rests on two 2026 arXiv preprints spanning two distinct lines of work. The raw field distribution is split evenly — 50% machine learning (efficient LLM inference) and 50% speech processing (audio-visual speech recognition). As first author, Park contributed ValueDiff, a value-geometric KV cache eviction method that removes tokens by their value-vector distance from the cache mean, reporting strong gains on sink-suppressed LLMs, though Evolve's review notes the absence of significance testing on the main results. As last author on the Korean lipreading paper, Park worked on personalized visual speech recognition on OLKAVS framed as transfer, census and adaptation, claiming Korean VSR state of the art at 9.95–12.19% CER versus 26.64, with LoRA-based speaker personalization and rigorous paired evaluation, albeit without prior-method baselines. The collaboration footprint is compact: 7 distinct co-authors, with Jungwook Choi, Mingu Lee, Se Un Park, Hakjun Kim and Taehoon Roh recurring, consistent with a small industry research group. Radar scores reflect this early, thin record — collaboration 41, reproducibility 55 (a mean over 2 papers, each rated medium), focus 0 from the even field split — while output, impact and activity axes remain unscored pending richer bibliometric data. Neither 2026 paper is marked corresponding-author.
Collaboration pattern
Collaboration runs through a compact, stable industry group: 7 distinct co-authors appear in the raw data, with Jungwook Choi, Mingu Lee, Se Un Park, Hakjun Kim and Taehoon Roh recurring across the 2026 preprints. Both papers place Junyoung Park in edge positions — first author on ValueDiff, last author on the OLKAVS lipreading work — and neither marks a corresponding author, suggesting senior-but-not-lead roles within Qualcomm's research lab rather than wide external networking.
Research keywords
Highlights
ValueDiff: value-geometric KV cache eviction
Personalized Korean lipreading on OLKAVS
First and last authorship within one year
In the 2026 catalog entries, this researcher appears as first author on the ValueDiff inference paper and last author on the OLKAVS speech paper — neither marked corresponding — a pattern consistent with senior technical contribution inside a small Qualcomm research group.
Timeline
- 2026Posts ValueDiff to arXiv as first author — value-geometric KV cache eviction for sink-suppressed LLMs, entering the efficient LLM inference line.
- 2026Publishes personalized Korean lipreading on OLKAVS as last author, claiming Korean VSR state of the art at 9.95–12.19% CER versus 26.64.
- 2026Both preprints enter the Evolve catalog with medium repro readiness; radar records collaboration 41, reproducibility 55 and an even 50/50 field split.
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.
- Bibliometrics come from a thin OpenAlex-derived snapshot: only 2 catalogued 2026 arXiv preprints; the output, impact and activity axes are unscored for lack of data.
- 'Junyoung Park' is a common Korean name; this profile is name-resolved from public bibliographic data and the listed Qualcomm (United States) affiliation, with no ORCID available for disambiguation.
- The reproducibility score (55) averages 2 papers, both rated medium repro readiness; verdicts reflect Evolve review, not independent replication.
- The frequent-coauthor list includes the profile subject's own name and one duplicate ('Se Un Park' appears twice); named collaborators here exclude the subject, and no country metadata was available.