Se-Young Yun

Research profile

Research summary

Se-Young Yun is a machine learning researcher at the Korea Advanced Institute of Science and Technology (KAIST), South Korea. This profile is name-resolved from public bibliographic data (OpenAlex; ORCID 0000-0001-6675-5113). The OpenAlex record reports an h-index of 20 and an i10-index of 42, with yearly output that accelerated after 2020 (15 works) and has stayed near its peak since — 28 works in 2022, 27 in 2023, 27 in 2024, and 24 in 2026 — totaling 69 works over the most recent three-year window, which drives a maximum activity score of 100. Topic tags span five clusters: domain adaptation and few-shot learning (58 works), advanced neural network applications (31), topic modeling (30), reinforcement learning in robotics (30), and advanced bandit algorithms (29), indicating a span from decision-making foundations to applied deep learning. The current field distribution splits evenly (0.5 and 0.5) between LLM post-training and efficient inference for long-context language models. Evolve's catalog holds three 2026 arXiv papers: on-policy reverse distillation for weak-to-strong generalization (last author; high reproducibility readiness), declarative attention control cutting attended tokens by 31–52% at a 1–3pp accuracy cost (middle author; medium readiness), and PLC-DPO for noisy preference optimization (last author; not yet reviewed). Two assessed papers average a reproducibility score of 78; collaboration scores 55 with 15 distinct recent coauthors.

Research keywords

Highlights

• On-policy reverse distillation elicits weak-to-strong generalization

  Evolve's review finds OPRD amplifies student RL gradients along a weak teacher's policy shift, with broad and solid results; single-seed runs and tuned references temper confidence. Reproducibility readiness: high, with Yun in the last author position.

• Language models learn to control their own attention

  The reviewed method (DA) lets LLMs declare their own attention zero-shot, cutting attended tokens by 31–52% at a 1–3pp accuracy cost; ablations are strong, though wall-clock gains are only projected. Reproducibility readiness: medium, with Yun as a middle author.

• PLC-DPO corrects noisy and ambiguous preference labels

  A 2026 arXiv paper on posterior label correction in noisy and ambiguous preference optimization, with Yun in the last author position. Evolve has not yet reviewed it — no verdict, keywords, or reproducibility assessment is available.

Timeline

1. 2008Earliest indexed year in the OpenAlex counts-by-year record, with 1 work — the start of the publication history shown in this snapshot.
2. 2020Indexed output accelerates to 15 works, up from 8 in 2019, opening a sustained productivity climb that continues through 2026.
3. 2022Peak indexed year at 28 works, with 27 following in both 2023 and 2024 — the core of the recent productivity surge.
4. 2026Three arXiv papers enter Evolve's catalog — weak-to-strong generalization via on-policy reverse distillation (last author), declarative self-controlled attention for long-context LLMs (middle author), and PLC-DPO for noisy preference optimization (last author) — alongside 24 indexed works this year.

Papers in the Evolve catalog

• 2026-09-08

  Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

• 2026-09-02

  Language Models Can Control Their Own Attention

Frequent co-authors