Jeonghye Kim
Microsoft Research Montréal (Canada) · South Korea
Microsoft Research Montréal researcher building verifiable SWE task pipelines and RL recipes for LLM coding agents.
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
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—
Not yet scored: the snapshot holds only two 2026 arXiv papers, too thin to measure publication volume, venue quality, or trajectory. This axis would track yearly output and publication venues as the record grows.
Impact—
Not yet scored: a two-paper 2026 record gives no citation signal (the snapshot reads h-index 0, i10-index 0). This axis would measure citations, uptake of the released benchmarks, and influence on downstream coding-agent research.
Focus0/100
The focus axis reads 0/100, the lowest measured score. The raw field distribution splits evenly (0.5/0.5) between AI for software engineering (coding agents) and LLM-agent reinforcement learning, and the two papers span four OpenAlex topics, from software engineering research to teaching and learning programming and testing and debugging. With only two papers, this even spread drives the score down, even though the strands are topically adjacent.
Collaboration60/100
Collaboration scores 60/100. The snapshot records 19 distinct coauthors, twelve of whom recur frequently — including Marc-Alexandre Côté, Alessandro Sordoni, Xingdi Yuan, and Zhengyan Shi — pointing to steady work inside a sizable Microsoft Research Montréal group rather than solo authorship. Both reviewed papers are multi-authored, and the single first-author position suggests Kim both leads a benchmark-infrastructure effort and contributes to larger agent-training projects.
Activity32/100
Activity scores 32/100, consistent with a just-starting record: countsByYear logs only 2 works, both in 2026, and recent3yWorks is likewise 2. Both papers appeared on arXiv within the same year, so the near-term signal is a concentrated burst of output rather than a sustained yearly cadence. This axis should be re-read after another publication cycle.
Reproducibility38/100
Reproducibility scores 38/100 (reproMean over the 2 reviewed papers). ProgramDistill is rated low readiness: strong engineering and replay verification of 4,063 tasks, but thin statistics and no public release yet. FrogNano is rated medium: the distillation-free RL recipe is credible, yet validation-set reuse, base-score discrepancies, and missing significance testing weaken independent replication.
Research summary
Jeonghye Kim is a researcher at Microsoft Research Montréal whose work sits at the intersection of AI for software engineering and machine learning for LLM agents. This profile is name-resolved from OpenAlex bibliographic data and rests on a very early record: two arXiv papers, both dated 2026, which together constitute the entire recent three-year output (recent3yWorks: 2). Kim is first author of ProgramDistill (2026), a pipeline that converts 26 interactive web apps into 4,063 replay-verified, reference-guided SWE tasks for coding agents, and a middle author of FrogNano (2026), which trains a Qwen3.5-4B coding agent via distillation-free reinforcement learning on roughly 1,500 online-synthesized tasks, reporting 61.5% on SWE-bench Verified. OpenAlex topics concentrate on Software Engineering Research and Software Engineering Techniques and Practices, with secondary ties to teaching and learning programming and to software testing and debugging; the field distribution splits evenly (0.5/0.5) between coding-agent AI and LLM-agent reinforcement learning. Collaboration is the strongest measured axis (60/100, 19 distinct coauthors); activity reads 32/100 and reproducibility 38/100, the latter reflecting thin statistics, no public release for ProgramDistill, and validation-set reuse in FrogNano.
Collaboration pattern
Kim works embedded in a multi-author Microsoft Research Montréal team: 19 distinct coauthors appear in the snapshot, and 12 recur frequently — including Marc-Alexandre Côté, Alessandro Sordoni, Xingdi Yuan, Zhengyan Shi, Young Jin Kim, and Minseon Kim. The collaboration dimension scores 60/100. The authorship pattern (first author on ProgramDistill, middle author on FrogNano) suggests Kim both leads benchmark-infrastructure work and contributes to larger agent-training projects within the same group.
Research keywords
Highlights
ProgramDistill: 26 web apps become 4,063 verifiable SWE tasks
FrogNano: a 4B coding agent trained on ~1,500 synthetic tasks
A closed loop from task synthesis to agent training
Read together, the two 2026 papers outline one coherent thread: synthesize verifiable tasks from real software, then train coding agents on them. This benchmark-plus-training agenda is reflected in the even 0.5/0.5 field split between coding-agent AI and LLM-agent reinforcement learning.
Timeline
- 2026Publishes ProgramDistill on arXiv as first author — an automated pipeline converting 26 interactive web apps into 4,063 replay-verified, reference-guided SWE tasks for coding agents.
- 2026Co-authors FrogNano (middle author), training a Qwen3.5-4B coding agent with distillation-free reinforcement learning on ~1,500 online-synthesized tasks, reporting 61.5% on SWE-bench Verified.
- 2026Both papers together form the entire record in this snapshot (countsByYear 2026: 2; recent3yWorks: 2), spanning coding-agent benchmarking and reinforcement-learning-based agent training.
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 and topics come from an OpenAlex name-resolved snapshot; 'Jeonghye Kim' is a common Korean name and no ORCID is linked, so identity resolution is probabilistic.
- Citation metrics are effectively absent: the snapshot reads h-index 0 and i10-index 0, so the output and impact axes are left unscored rather than estimated.
- Only two arXiv papers (both 2026) are reviewed; the data contains no peer-reviewed venue, award, or career-history information.
- The reproducibility score (38/100) reflects Evolve's readiness review of just 2 papers, not independent replication attempts.