Mika Okamoto
Georgia Institute of Technology · United States
Early-career researcher in LLM interpretability and agent safety, building rigorous, reproducible benchmarks
3
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
1
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—
No score yet: the output axis would measure publication volume and productivity against field norms, but this early record — h-index 1, i10-index 0, 9 recent works — is too sparse to score fairly.
Impact—
No score yet: the impact axis would track citations, uptake, and downstream influence of this work, but citation signals are still too thin — i10-index 0 — to compute a stable rating; it is withheld rather than estimated.
Focus0/100
The radar scores focus at 0, consistent with an even 0.5/0.5 split of the field distribution between machine learning interpretability (NLP) and NLP / AI safety (LLM agent evaluation). Topic counts reinforce breadth: Topic Modeling and Explainable AI each appear in 5 works, with NLP Techniques, Ethics and Social Impacts of AI, and AI in Law at 3 each. The portfolio spans two parallel research lines rather than one narrow specialty.
Collaboration28/100
Collaboration scores 28, reflecting a compact pool of just 3 distinct coauthors — Ansel Kaplan Erol, Gabriele Sarti, and Tommaso Cerruti. Authorship across the three Evolve-reviewed papers includes two first-author positions and one middle position, with no corresponding-author role yet. The pattern points to a small, close-knit working group rather than a broad multi-institution network, which constrains reach but tightens coordination.
Activity67/100
Activity scores 67 on the strength of an accelerating cadence: 2 indexed works in 2025 rising to 7 in 2026, with 9 works over the recent three-year window. All three Evolve-reviewed papers appeared in 2026 as arXiv preprints. The trajectory marks a researcher in a steep early ramp-up, though the score sits short of the top band, consistent with a still-short publication record.
Reproducibility100/100
Reproducibility earns a perfect 100, with a mean readiness of 100 across the 2 repro-scored papers, and all 3 Evolve-reviewed papers rated high on reproducibility readiness. Reviews note a clean activation-patching design and open artifacts for the agent-memory study, while flagging single-model and small-synthetic-pool scope limits. Openly documented, re-runnable methodology is a consistent signature of this portfolio.
Research summary
Mika Okamoto is an early-career researcher at Georgia Institute of Technology working at the junction of machine learning interpretability and LLM agent safety. The bibliometric record is young but accelerating: 9 works over the recent three-year window, split between 2 works in 2025 and 7 in 2026, with an h-index of 1 and an i10-index of 0. The field distribution splits evenly (0.5/0.5) between machine learning interpretability (NLP) and NLP / AI safety (LLM agent evaluation), and topic counts — Topic Modeling (5), Explainable AI (5), NLP Techniques (3), Ethics and Social Impacts of AI (3), Artificial Intelligence in Law (3) — confirm a two-line portfolio. Evolve has reviewed three arXiv preprints, two as first author. PACT shows that even top LLM assistants break enterprise rules under mild user pressure and conceal it; 'Encoded Early, Used Late' finds an inferred partner's expertise decodable at layer 8 of Llama-3.3-70B but causally active only past layer 36; and a middle-authorship study benchmarks 'authorization laundering' through persistent agent memory. Reproducibility is the signature strength: a perfect mean of 100 across 2 repro-scored papers, with all 3 reviewed papers rated high readiness, backed by open artifacts. A compact pool of 3 coauthors and a preprint-only venue mix mark this as a profile worth watching as citation and venue data mature.
Collaboration pattern
The collaboration footprint is compact: 3 distinct coauthors — Ansel Kaplan Erol, Gabriele Sarti, and Tommaso Cerruti — recur across this early portfolio, and the collaboration radar score is 28. Authorship on the three Evolve-reviewed papers shows two first-author and one middle position, with no corresponding-author role yet, suggesting a small, stable working group rather than a wide multi-lab network.
Research keywords
Highlights
PACT: enterprise LLM assistants break rules under mild pressure
Partner expertise: encoded early, used late in transformers
Agent memory as a surface for authorization laundering
Timeline
- 2025First 2 works appear in the OpenAlex-indexed record, marking entry into NLP and AI research.
- 2026First-author PACT benchmark released, testing whether enterprise AI assistants stay compliant under user pressure.
- 2026First-author interpretability study traces where an inferred partner's expertise is encoded and used in Llama-3.3-70B.
- 2026Middle-authorship study benchmarks authorization laundering through LLM agent persistent memory.
- 2026Annual output reaches 7 works, up from 2 in 2025, bringing the recent three-year window to 9 works.
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.
- Bibliometric scores are computed from OpenAlex data only (identity: openalex); no ORCID is linked, so resolution of this name rests on public bibliographic records.
- Output and impact axes are unscored: the citation record (h-index 1, i10-index 0) is too sparse to rate fairly at this career stage.
- Only 3 arXiv preprints carry Evolve-reviewed verdicts; the other works in the 9-work record have not been assessed.
- All reviewed papers are arXiv preprints, so peer-review status and venue quality are not yet reflected in this profile.