NLP researcher at Chung-Ang University: decoding-time hallucination mitigation and safety steering for LLMs
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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: this snapshot lacks the underlying output data on total work counts and productivity over time. This axis would measure the volume and consistency of the publication record; it is left unscored rather than estimated.
Impact—
Not yet scored: citation data is insufficient — the snapshot reports h-index 0 and i10-index 0 with no per-paper citation counts. This axis would measure scholarly influence through citations and derived indices; it is left unscored rather than estimated.
Focus100/100
Focus scores a maximal 100/100. The raw field distribution lists exactly one field — natural language processing (LLM hallucination mitigation) — and OpenAlex topic counts reinforce the concentration: Natural Language Processing Techniques (6) and Topic Modeling (6) sit alongside Multimodal Machine Learning Applications (10), showing a profile channeled almost entirely into language-model intervention research rather than spread across many areas.
Collaboration41/100
Collaboration scores 41/100, derived from 7 distinct coauthors in the snapshot. The recurring partner list — Hoejoon Kwon, Kahyeon Kim, YoungBin Kim, Hayeong Ryu, JungMin Yun and Sunhee Jo — suggests a compact, single-team working pattern. Both catalogued 2026 papers place the researcher in a middle authorship position rather than as corresponding author, consistent with an early-career contributor inside an established group.
Activity79/100
Activity scores 79/100. OpenAlex records 14 works in the trailing three-year window (recent3yWorks: 14), and the counts-by-year series places all 14 in 2026 — a concentrated output burst at the very start of the record. Whether this pace is sustained across multiple years cannot yet be verified from a snapshot this young.
Reproducibility55/100
Reproducibility scores 55/100, computed as a mean over just one reviewed paper (reproPapers: 1). That paper, the factual-decoding study, carries a 'medium' Evolve repro-readiness rating, and its review notes missing significance testing. The score should be read as a single-paper signal, not a validated track record.
Research summary
Byeonggeuk Lim is an early-stage researcher at Chung-Ang University in South Korea whose public bibliographic record, resolved via OpenAlex, centers on natural language processing for large language models — specifically hallucination mitigation and safety steering. The snapshot is young and compact: 14 works appear in the trailing three-year window, all dated 2026, with h-index 0 and i10-index 0, consistent with a profile just entering the literature. Two arXiv preprints are catalogued. 'Look Before You Leap: Factual Decoding with Internal Attribution Signals' (2026) intervenes at decoding time using internal attribution signals grounded in causal tracing, MLP ablation and probing classifiers; Evolve's review credits factuality gains at 1.10× overhead while flagging weak short-answer QA and missing significance testing, and rates repro-readiness medium. 'ALTSTEER' (2026) explores selective safety steering that replaces hard refusals with constructive alternatives and awaits review. Topic counts reinforce the concentration: Multimodal Machine Learning Applications (10), Domain Adaptation and Few-Shot Learning (8), Topic Modeling (6), Natural Language Processing Techniques (6) and Privacy-Preserving Technologies in Data (4). The radar shows maximal focus (100), strong activity (79), moderate collaboration (41 across 7 distinct coauthors) and a single-paper reproducibility signal (55) — an emerging, tightly scoped profile whose citation footprint has yet to form.
Collaboration pattern
Collaboration scores 41/100 on a compact network of 7 distinct coauthors. Recurring partners are Hoejoon Kwon, Kahyeon Kim, YoungBin Kim, Hayeong Ryu, JungMin Yun and Sunhee Jo (one name repeats in the frequency list), a pattern consistent with a single lab or project team. Both 2026 catalog papers place the researcher in a middle authorship position rather than as corresponding author, indicating an early-career contributor role within an established group.
Research keywords
Highlights
Decoding-time factual decoding guided by internal attribution signals
Selective safety steering beyond hard refusals
The ALTSTEER preprint (2026, arXiv) targets selective safety steering that replaces blanket hard refusals with constructive alternatives. It is in the Evolve catalog but has not yet received a review verdict, keywords or repro-readiness rating.
Concentrated first-year output with maximal topical focus
The snapshot records 14 works in the trailing three-year window, all dated 2026, with a single listed field (natural language processing — LLM hallucination mitigation) and a focus score of 100 — a tightly scoped, fast-starting research agenda.
Timeline
- 2026arXiv preprint 'Look Before You Leap: Factual Decoding with Internal Attribution Signals' enters the Evolve catalog, reviewed with factuality gains at 1.10× overhead and a medium repro-readiness rating.
- 2026Second arXiv preprint, ALTSTEER, proposes selective safety steering that moves beyond hard refusals toward constructive alternatives; it awaits Evolve review.
- 2026OpenAlex snapshot records 14 works in the trailing three-year window — all dated 2026 — with h-index 0 and i10-index 0.
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 an OpenAlex name-resolved snapshot (identity: openalex); common-name ambiguity cannot be fully excluded.
- Output and impact are unscored: the snapshot reports h-index 0 and i10-index 0 with no per-paper citation counts.
- Only one catalog paper has a full Evolve review and reproducibility rating (reproMean 55 from 1 paper); ALTSTEER lacks a verdict, keywords and slug.
- Venue information beyond arXiv source labels is absent, so peer-reviewed publication status cannot be assessed.