Gyeongsik Yang

Research profile

Research summary

Gyeongsik Yang is a systems researcher at Korea University whose record spans cloud and network infrastructure and, most recently, the systems layer of large language models. OpenAlex topics trace a broad base — cloud computing and resource management (41 works), software-defined networks and 5G (30), and advanced optical networks (19) — now paired with an even 50/50 split between distributed LLM training and computer systems performance evaluation. In 2026, three last-authored arXiv papers entered the Evolve catalog: CONA switches parallelism strategies online for 1.4–9.6× faster time-to-perplexity on models up to 1.3B parameters; PARSER selects residual-pruning targets for mixture-of-experts LLMs by output importance, narrowing accuracy gaps against the state of the art at equal memory; and a characterization study shows AI-agent resource bottlenecks are task-dependent, demonstrated through admission control and allocation. The bibliometric snapshot reports an h-index of 12, an i10-index of 18, 25 works in the last three years, and 13 works already recorded in 2026 — an activity profile rated 95/100. Collaboration is moderate (55/100), with 15 distinct co-authors in the recent window and a recurring cluster around Chuck Yoo, plus Zhixiong Niu and Yongqiang Xiong. Reproducibility is rated 55/100; reviewed papers are single-host or single-run in places, and the impact and output axes are not yet scored.

Research keywords

Highlights

• CONA: parallelism strategy chaining cuts training time-to-perplexity by 1.4–9.6×

  Evolve's review credits CONA with switching parallelism strategies online for 1.4–9.6× faster time-to-perplexity; results are consistent but small-scale, single-run, and confined to models of at most 1.3B parameters — a practical lever for training-efficiency work whose generality is still to be proven.

• PARSER: output-importance-guided residual sparsification for MoE LLM compression

  Reviewed as selecting MoE residual-pruning targets by output importance, PARSER narrows accuracy gaps against the state of the art at equal memory; the review notes gains skew MMLU, while reproducibility readiness is rated high — the strongest replication signal among Yang's cataloged papers.

• AI-agent resource bottlenecks shown to be task-dependent

  A careful characterization of agent tool-container workloads finds resource and performance dynamics vary by task, with admission-control and allocation demonstrations that convince — though they rest on one host, one LLM, and replayed LLM trajectories, per the Evolve verdict.

Timeline

1. 2016Earliest year in the OpenAlex counts-by-year record: 3 works recorded.
2. 2020Annual output reaches 6 works, the level sustained through 2022, within the cloud-and-network era that dominates the topic profile (cloud computing 41, SDN/5G 30, optical networks 19).
3. 2023Dips to 3 works — the lowest annual count since 2016 — before the current climb.
4. 2025Returns to 6 works, setting up the 2026 surge.
5. 202613 works recorded so far; three last-authored arXiv papers enter the Evolve catalog — CONA (parallelism strategy chaining), PARSER (MoE residual sparsification), and the AI-agent resource characterization study.

Papers in the Evolve catalog

• 2026-09-18

  Not All AI Agents Are Equal: Characterizing Resource and Performance Dynamics

• 2026-09-07

  Parallelism Strategy Chaining for Fast Training Convergence

• 2026-09-01

  Residual Sparsification via Output Importance for Compressing Mixture-of-Experts LLMs

Frequent co-authors