Seogyeong Jeong (KAIST)

Korea Advanced Institute of Science and Technology · South Korea

Unclaimed · AI-structured profileOpenAlex

KAIST researcher in LLM mechanistic interpretability and safety; first-authored work on chain-of-thought reasoning

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Flagged missing: the snapshot reports h-index 0 and i10-index 0, so no stable volume score exists yet. This axis would measure cumulative works, venue spread, and first/corresponding-author share across the career.

Impact—

Flagged missing: with h-index 0 and i10-index 0 in the OpenAlex snapshot, citation-based scoring is not yet meaningful. This axis would measure citations, citing works, and influence on follow-up research.

Focus0/100

The radar assigns focus a score of 0, the lowest on this profile. The raw data pulls in two directions: the reviewed-work field split is 0.5/0.5 across two closely related mechanistic-interpretability areas, while OpenAlex topic tags scatter across speech and dialogue systems (4), topic modeling (3), speech recognition and synthesis (3), NLP techniques (2), and AI ethics (2). A fair reading is an early-stage profile still consolidating around LLM interpretability.

Collaboration41/100

Collaboration scores 41 on the 0-100 radar. The snapshot records 7 distinct coauthors over the last three years alongside 8 recent works, with a recurring KAIST-centered cluster: Alice Oh, Soyeon Park, and Sunwoo Kim appear as frequent coauthors, alongside Jaehui Hwang, Dongyoon Han, Geonmo Gu, and Taekyung Kim. Authorship spans first author on the chain-of-thought study and middle author on the harm-signals paper, neither corresponding — a co-authorial, lab-anchored pattern.

Activity64/100

Activity scores 64, the strongest axis on this profile. OpenAlex counts 8 works in the last three years, with counts-by-year rising from 3 in 2025 to 5 in 2026 — an accelerating ramp rather than a fading tail. Both cataloged papers are 2026 arXiv postings, one first-authored and one middle-authored, indicating a currently productive phase concentrated in LLM interpretability and safety.

Reproducibility55/100

Reproducibility scores 55, computed by Evolve from 2 reviewed papers (reproMean 55), each rated 'medium' for repro-readiness. Verdicts credit solid empirical control — the chain-of-thought study shows separable middle-layer geometry well controlled for confounds — while flagging thin statistics and GPT-5 label noise with split leakage as limits on how cleanly the results would re-run.

Research summary

Seogyeong Jeong is a KAIST researcher working at the intersection of natural language processing and the mechanistic interpretability of large language models, with a focus on LLM safety. The Evolve catalog holds two 2026 arXiv papers: a first-authored mechanistic interpretation of chain-of-thought reasoning operations, which finds that reasoning operations form separable middle-layer geometry under careful confound control, and a middle-authored study of fine-grained harm signals showing that orthogonal category-specific harm directions still carry harmfulness and can amplify downstream general-harm alignment. Both carry 'medium' repro-readiness ratings, with an aggregate reproducibility score of 55 across the 2 reviewed papers. Bibliometric coverage comes from OpenAlex: h-index and i10-index are both 0, output and impact axes are therefore flagged missing, and the profile should be read as early-stage. Activity is the strongest signal (64), with 8 works in the last three years and counts-by-year rising from 3 in 2025 to 5 in 2026. Collaboration scores 41, anchored in a KAIST cluster — Alice Oh, Soyeon Park, and Sunwoo Kim — plus Jaehui Hwang, Dongyoon Han, Geonmo Gu, and Taekyung Kim, across 7 distinct coauthors. Focus scores 0: topic tags scatter across speech and dialogue systems, topic modeling, and speech recognition while the field split is even between two mechanistic-interpretability areas, so topical consolidation is still in progress.

Collaboration pattern

Collaboration is KAIST-anchored and mid-range (score 41): 7 distinct coauthors across the last three years. The recurring cluster includes Alice Oh, Soyeon Park, and Sunwoo Kim (all KAIST), plus Jaehui Hwang, Dongyoon Han, Geonmo Gu, and Taekyung Kim. The pattern pairs first-author leadership on the chain-of-thought mechanistic study with middle-authorship on the harm-signals paper; neither cataloged paper flags corresponding authorship, consistent with a lab-based early-career collaboration style.

Research keywords

Mechanistic interpretabilityLLM safetyActivation steeringChain-of-thought reasoningLinear probingRepresentation geometryRepresentation engineeringRefusal behaviorHarmfulness representationsReasoning operationsAttention maskingEthics and social impacts of AI

Highlights

First-authored mechanistic account of chain-of-thought reasoning
Beneath the Surface of Chains-of-Thought (2026, arXiv, first author) finds that CoT reasoning operations form separable middle-layer geometry under careful confound control; Evolve notes GPT-5 label noise and split leakage temper the mechanistic claims.
Fine-grained harm directions in LLM safety
The Role of Fine-grained Harm Signals in LLM Safety (2026, arXiv, middle author) shows orthogonal category-specific harm directions still carry harmfulness and amplify downstream general-harm alignment; Evolve calls it a solid empirical study with thin statistics.
Medium repro-readiness across both cataloged studies
Evolve rates both papers 'medium' for reproducibility readiness; the profile-level reproducibility score is 55 over the 2 reviewed papers, backed by confound controls but limited by thin statistics and label noise.

Timeline

2025OpenAlex counts-by-year records 3 works, the earliest year covered in this profile's snapshot.
2026First-authored 'Beneath the Surface of Chains-of-Thought' posted to arXiv, mechanistically probing reasoning operations in LLMs.
2026Co-authored 'The Role of Fine-grained Harm Signals in LLM Safety' posted to arXiv, extending the safety line to category-level harm directions.
2026OpenAlex counts 5 works for the year — the highest annual count in this snapshot — within 8 works over the last three years.

Papers in the Evolve catalog

2026-09-18
The Role of Fine-grained Harm Signals in LLM Safety
2026-09-04first author
Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs

Frequent co-authors

Alice Oh (KAIST)×2 Soyeon Park (KAIST)×1 Sunwoo Kim (KAIST)×1 Jaehui Hwang×1 Dongyoon Han×1 Geonmo Gu×1 Taekyung Kim×1

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.

All bibliometrics come from an OpenAlex snapshot (identity: openalex); no ORCID was provided, so the profile is name-resolved from public bibliographic data.
h-index and i10-index are 0 in the snapshot, and the output and impact dimensions are flagged missing — citation-based claims are not yet meaningful.
Evolve has reviewed only 2 papers; the qualitative verdicts cover this small subset and may not represent the full body of work.
The reproducibility score of 55 rests on 2 papers rated 'medium', a thin base for a stable rating.