Alice Oh (KAIST)

Korea Advanced Institute of Science and Technology · Denmark

Unclaimed · AI-structured profileOpenAlex

KAIST-based NLP researcher focused on mechanistic interpretability and safety of large language models

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—

No aggregate output score is available: the snapshot reports per-year work counts but no total-volume measure for this record. This axis would capture overall publication productivity — number of works over time and across venues — so it is left unscored rather than estimated.

Impact—

No impact score is available: the record reports only an h-index of 1 and an i10-index of 0, too thin to score citation influence reliably. This axis would measure citation-based impact — total citations, h-index and citation trajectory — and remains unscored.

Focus0/100

The focus score of 0 reflects strong dispersion rather than concentration: the record's field distribution splits evenly (0.5/0.5) between mechanistic interpretability of LLM safety and of large language models, while its five OpenAlex topics span topic modeling (14), NLP techniques (9), hate speech and cyberbullying detection (8), ethics and social impacts of AI (6) and adversarial robustness (5). The two 2026 catalog papers do converge on interpretability and safety, so the dispersion partly reflects the record's earlier history.

Collaboration41/100

A collaboration score of 41 sits in the middle range: over the last three years the record lists 24 works with 7 distinct coauthors, a small and stable pool. The frequent-coauthor list is largely KAIST-affiliated — Seogyeong Jeong, Soyeon Park and Sunwoo Kim (KAIST), plus Jaehui Hwang, Dongyoon Han, Geonmo Gu and Taekyung Kim — indicating a primarily institution-centered network rather than wide external collaboration.

Activity94/100

An activity score of 94 reflects a steep recent ramp: OpenAlex indexes 2 works in 2024, 3 in 2025 and 19 in 2026, totalling 24 works over the last three years. The large majority of this output concentrates in 2026, the year both Evolve-reviewed arXiv papers appeared.

Reproducibility55/100

A reproducibility score of 55 is the mean across the 2 papers Evolve has assessed (reproMean 55, reproPapers 2). Both catalog papers — on fine-grained harm signals and on chain-of-thought reasoning operations — are rated medium for reproducibility readiness, with review verdicts noting thin statistics in one and GPT-5 label noise plus split leakage in the other.

Research summary

Alice Oh is a researcher affiliated with the Korea Advanced Institute of Science and Technology (KAIST) whose recent work centers on the mechanistic interpretability and safety of large language models. The Evolve catalog includes two 2026 arXiv papers: a study of fine-grained harm signals in LLM safety, showing that category-specific harm directions retain harmfulness and amplify downstream general-harm alignment (last author), and a mechanistic interpretation of chain-of-thought reasoning operations, finding separable middle-layer geometry (middle author). Both are rated medium for reproducibility readiness, with an average reproducibility score of 55 across 2 assessed papers. OpenAlex signals point to sharply accelerating output: 2 works indexed in 2024, 3 in 2025, and 19 in 2026, totalling 24 works in the last three years with 7 distinct coauthors, yielding an activity score of 94. The record's topic distribution is unusually broad — topic modeling (14), NLP techniques (9), hate speech and cyberbullying detection (8), ethics and social impacts of AI (6), and adversarial robustness (5) — which, together with a reported h-index of 1 and i10-index of 0, indicates a name-resolved profile that may be split or conflated across records. Aggregate output and impact scores are unavailable in the snapshot and remain unscored.

Collaboration pattern

Collaboration is concentrated and institution-centered: 7 distinct coauthors across the last three years (collaboration score 41), with the frequent-coauthor list dominated by KAIST colleagues — Seogyeong Jeong, Soyeon Park and Sunwoo Kim (KAIST), plus Jaehui Hwang, Dongyoon Han, Geonmo Gu and Taekyung Kim. On the two Evolve-reviewed 2026 papers the researcher is last author (non-corresponding) on one and middle author on the other, a pattern often read as a senior authorship role, though correspondence is not flagged.

Research keywords

mechanistic interpretabilityLLM safetyactivation steeringchain-of-thought reasoningrepresentation engineeringrefusal behaviorlinear probingtopic modelinghate speech detectionAI ethics and social impactadversarial robustnessnatural language processing

Highlights

Fine-grained harm directions stay harmful and reshape general alignment
Evolve-reviewed 2026 arXiv study (last author) finds that category-specific harm directions in LLMs still carry harmfulness and amplify downstream general-harm alignment; the verdict calls it a solid empirical study with thin statistics. Reproducibility readiness: medium.
Chain-of-thought reasoning leaves separable middle-layer geometry
2026 arXiv mechanistic interpretation of reasoning operations (middle author) shows CoT operations form separable middle-layer geometry, well controlled for confounds; GPT-5 label noise and split leakage temper the mechanistic claims. Reproducibility readiness: medium.
Sharp output acceleration into 2026
OpenAlex indexes 19 works for this record in 2026, up from 3 in 2025 and 2 in 2024; both Evolve-reviewed mechanistic-interpretability papers fall within this 2026 cluster, marking a pivot toward LLM safety and interpretability.

Timeline

20242 works indexed in OpenAlex for this record — the earliest year in the snapshot's counts-by-year data.
20253 works indexed; output remained modest before the 2026 acceleration.
202619 works indexed, including both Evolve-reviewed arXiv papers: fine-grained harm signals in LLM safety (last author) and mechanistic interpretation of chain-of-thought reasoning (middle author).

Papers in the Evolve catalog

Frequent co-authors

Seogyeong Jeong (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.

Bibliometrics are computed from OpenAlex only; no ORCID is provided and the profile is name-resolved from public bibliographic data.
Output and impact axes are unscored because the snapshot lacks aggregate publication and citation metrics; the reported h-index of 1 and i10-index of 0 are the only citation figures available.
The record's topic history (topic modeling, hate speech detection, AI ethics) differs markedly from the recent mechanistic-interpretability cluster, so the OpenAlex identity may conflate or split records; verify before citing.
The affiliation country is recorded as 'DK', which is inconsistent with the KAIST affiliation; institutional metadata should be verified.