Manabu Okumura

Institute of Science Tokyo · Japan

Unclaimed · AI-structured profileORCID OpenAlex

Science Tokyo NLP researcher: four decades of text analysis, now on LLM watermarking, personalization, and MT decoding.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Total publication-volume metrics are not yet available in this snapshot, so no output score is shown. This axis would measure the cumulative number of indexed works and how that volume is distributed across years and publication venues.

Impact—

Citation-impact scoring is not yet computed. The snapshot does record an h-index of 32 and an i10-index of 97, but without a per-paper citation distribution this axis — which would measure citation counts, field percentiles, and influence — stays unscored.

Focus0/100

The focus score is 0, reflecting breadth rather than concentration: the recorded field distribution splits evenly at 0.5/0.5 between LLM text watermarking and LLM personalization, and indexed topics span topic modeling (267), NLP techniques (229), advanced text analysis (110), sentiment analysis and opinion mining (71), and information retrieval and search behavior (69). This is a wide portfolio rather than a single-thread agenda.

Collaboration56/100

Co-authorship breadth is moderate (score 56): 16 distinct co-authors appear in the recent three-year window. Recurring partners include Hidetaka Kamigaito, Kotaro Funakoshi, and Taro Watanabe, alongside Guangyu Chen, Boxuan Lyu, Sangjun Moon, Dasom Choi, and Jingun Kwon on the 2026 work. Authorship positions on the reviewed papers run from last author (twice) to middle author (once).

Activity100/100

Activity is the strongest axis (score 100): the snapshot records 98 works in the last three years, with per-year indexed counts moving from 29 (2023) to 39 (2024), 21 (2025), and 38 (2026) — sustained high-volume output more than four decades after the earliest indexed entries in 1986.

Reproducibility100/100

Reproducibility is scored 100, with a mean reproducibility metric of 100 computed across 2 assessed papers. Within the Evolve catalog, TripPattern and HyperTrace carry high reproducibility readiness and Quit While You're Ahead is rated medium, consistent with method write-ups that specify procedures such as the runs test and TOST equivalence testing.

Research summary

Manabu Okumura is a natural language processing researcher at the Institute of Science Tokyo whose indexed record spans 1986–2026, with an h-index of 32 and an i10-index of 97 in the OpenAlex-derived snapshot. His topic profile is anchored in topic modeling (267 indexed assignments), NLP techniques (229), advanced text analysis (110), sentiment analysis and opinion mining (71), and information retrieval and search behavior (69), reflecting career-long breadth across text analytics. Recent activity is high: 98 works in the last three years, with per-year indexed output peaking at 39 (2024) and 38 (2026). His current work splits evenly between LLM text watermarking and LLM personalization (0.5/0.5 field distribution). The three Evolve-reviewed 2026 arXiv preprints show this range: TripPattern proposes pattern-based, three-group vocabulary-partition watermarking for LLMs; HyperTrace applies training-free Sequential Monte Carlo hypothesis tracing to online LLM personalization; and Quit While You're Ahead introduces early stopping for candidate generation in machine translation reranking, reporting 1.47–4.12× speedups. He is last author on two of these and middle author on one, with 16 distinct co-authors in the recent window. Reproducibility signals are strong, with a mean reproducibility score of 100 across the two assessed papers.

Collaboration pattern

Co-authorship is moderately broad: 16 distinct co-authors in the recent three-year window (collaboration score 56). Recurring partners include Hidetaka Kamigaito, Kotaro Funakoshi, and Taro Watanabe, with additional collaborators on the 2026 arXiv work such as Guangyu Chen, Boxuan Lyu, Sangjun Moon, Dasom Choi, Jingun Kwon, Jianzhi Shen, Keyu Mao, Minghao Shao, and Chuanyang Jin. On the reviewed papers he takes senior positions — last author twice and middle author once.

Research keywords

LLM text watermarkingmachine-generated text detectionLLM personalizationonline preference learningtopic modelingsentiment analysis and opinion mininginformation retrievalmachine translationMBR decodinginference efficiencySequential Monte Carlotraining-free adaptation

Highlights

TripPattern: pattern-based text watermarking that preserves LLM output quality
As last author, he contributes to a watermarking scheme that alternates three vocabulary groups instead of KGW-style green/red lists, validated with the Wald–Wolfowitz runs test. Evolve's verdict: it preserves text quality better than KGW with broad multilingual results, though it remains weak against paraphrase attacks. Reproducibility readiness: high.
HyperTrace: training-free hypothesis tracing for online LLM personalization
On this middle-author 2026 preprint, the method frames online preference learning as tracing natural-language user hypotheses with Sequential Monte Carlo, combined with a Bradley-Terry choice model and requiring no training. Evolve notes the effects appear consistent but moderate, judged mainly by LLM evaluators on small samples. Reproducibility readiness: high.
QUIT: early stopping that speeds up machine translation reranking
As last author, he co-authors a method that halts candidate generation once the best reranking score stabilizes, achieving 1.47–4.12× speedups with translation quality verified by TOST equivalence. Evolve rates the work solid while noting that quality-estimation-based stopping nearly degenerates. Reproducibility readiness: medium.

Timeline

2026All three Evolve-reviewed works appear in this single year, within a three-year run of 98 indexed works (38 indexed in 2026 itself).
2026TripPattern preprint introduces pattern-based, three-group vocabulary-partition watermarking for large language models.
2026HyperTrace preprint applies Sequential Monte Carlo hypothesis tracing to online LLM personalization, training-free.
2026QUIT preprint adds stability-based early stopping to machine translation reranking, with 1.47–4.12× speedups.

Papers in the Evolve catalog

Frequent co-authors

Hidetaka Kamigaito×2 Kotaro Funakoshi×2 Guangyu Chen×1 Boxuan Lyu×1 Sangjun Moon×1 Dasom Choi×1 Jingun Kwon×1 Taro Watanabe×1 Jianzhi Shen×1 Keyu Mao×1 Minghao Shao×1 Chuanyang Jin×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.

Bibliometric indicators (h-index, i10-index, per-year counts, topics) come from an OpenAlex-derived snapshot; output and impact dimension scores are not yet computed.
Only the three 2026 arXiv preprints reviewed by Evolve are individually assessed; other indexed works are not analyzed.
Identity is name-resolved from public bibliographic data linked to ORCID 0009-0001-7730-1536; no biographical information beyond the snapshot is used.
The reproducibility mean of 100 is based on only 2 papers with computed metrics, so the axis rests on a small sample.