YoungBin Kim

Research profile

Research summary

YoungBin Kim is a researcher at Chung-Ang University whose five catalogued 2026 arXiv preprints sit at the intersection of natural language processing, LLM safety alignment, and information retrieval — a spread the bibliometric field distribution weights almost exactly evenly (roughly one-third each). Kim appears as the final author on all catalogued papers, a pattern consistent with a senior or supervisory role, though none is marked corresponding. Three lines of work stand out. First, reward-based, outcome-only evaluation for grammatical error correction: a DeBERTa reward model trained on LLM-judged GEC preferences, with strong SEEDA results. Second, segment-aware listwise direct preference optimization for reasoning safety in large reasoning models. Third, training-free multi-hop retrieval built on contrastive facet indexing, sequential evidence exploration, and coverage-aware ranking. Two further 2026 preprints extend the safety agenda — ALTSTEER's selective steering beyond hard refusals toward constructive alternatives, and an audit of dialect penalties in text-to-image safety pipelines. Across the three papers Evolve has reviewed, reproducibility averages 55/100 (medium readiness); reviewers credit strong reported gains but flag reliance on proprietary or same-family LLM judges and missing significance tests. The profile is name-resolved from public bibliographic data; output, impact, and activity metrics are not yet available.

Research keywords

Highlights

• Outcome-only, reward-based evaluation for grammatical error correction

  The paper trains a DeBERTa reward model on LLM-judged GEC preferences and evaluates by outcome rather than edit counts. Evolve finds the SEEDA results strong, but notes the supervision rests on proprietary LLM judges with limited rigor.

• Segment-aware listwise alignment for reasoning safety in large reasoning models

  The work proposes segment-aware listwise DPO to align chain-of-thought safety in large reasoning models. Evolve judges the strong gains credible, but flags reliance on same-family judge scoring and the absence of significance tests.

• Training-free contrastive evidence exploration for multi-hop retrieval

  A training-free multi-hop retrieval method combining contrastive facet indexing, sequential evidence exploration, and coverage-aware ranking. Evolve reports strong gains, but notes missing significance tests and insufficient corpus details.

• A broadening safety agenda: steering beyond refusals and auditing dialect bias

  Two further 2026 preprints — ALTSTEER on selective safety steering toward constructive alternatives, and an audit of dialect penalties in text-to-image safety pipelines — extend the group's safety focus beyond the reviewed papers. Neither has yet been reviewed by Evolve.

Timeline

1. 2026Publishes a reward-based, outcome-only evaluation framework for grammatical error correction, training a DeBERTa reward model on LLM-judged preferences; Evolve rates the SEEDA results strong.
2. 2026Proposes segment-aware listwise DPO for reasoning safety in large reasoning models, moving beyond single-turn safe answers toward chain-of-thought safety.
3. 2026Introduces a training-free multi-hop retrieval approach built on contrastive facet indexing, sequential evidence exploration, and coverage-aware ranking.
4. 2026Releases ALTSTEER, a selective safety-steering method that moves beyond hard refusals toward constructive alternatives (not yet reviewed by Evolve).
5. 2026Audits the dialect penalty in text-to-image safety pipelines, showing such pipelines are not safe for all users (not yet reviewed by Evolve).

Papers in the Evolve catalog

• 2026-09-14

  Don't Count the Edits, Judge by the Outcome Alone: Reward-Based Evaluation for Grammatical Error Correction

• 2026-09-14

  Beyond Safe Answers: Segment-Aware Listwise Alignment for Reasoning Safety in Large Reasoning Models

• 2026-09-07

  Beyond One-Shot Expansion: Contrastive Evidence Exploration for Multi-Hop Retrieval

Frequent co-authors