Gouki Minegishi

The University of Tokyo

Unclaimed · AI-structured profileOpenAlex

University of Tokyo researcher probing in-context world representations and self-monitoring in large language models.

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—

Publication-output metrics such as paper counts, output volume and venue spread over time are not yet computed for this profile, so no score is shown; current verifiable evidence rests on two catalogued 2026 arXiv preprints.

Impact—

Impact measures such as citations, downloads and downstream uptake are unavailable in the current data, so this axis is unscored; both catalogued papers are recent 2026 preprints whose citation records have not yet accumulated.

Focus100/100

Evolve's focus dimension scores 100/100, and the raw field distribution explains why: all catalogued output falls in a single bucket, machine learning (LLM interpretability and in-context learning). Both reviewed papers are interpretability studies of large language models — one on in-context world representations, one on privileged-access neurofeedback — sharing probing and causal-intervention methodology. This is unusually tight concentration, though it rests on only two 2026 preprints, so the score should be read as directional rather than settled.

Collaboration50/100

The collaboration dimension scores 50/100, based on 11 distinct co-authors recorded across the catalogued work. The provided frequent-coauthor list is Tokyo-centric and repeats names: Yutaka Matsuo, Yusuke Iwasawa and Takeshi Kojima each appear twice, joined by Kohsei Matsutani, Core Francisco Park, Koshiro Aoki, Ryota Takatsuki, Yusuke Haruki and Daisuke Kawahara. On both papers Minegishi holds a middle, non-corresponding position, fitting a contributor role within a larger interpretability team rather than a solo or leading-authorship pattern.

Activity—

Activity metrics such as publication cadence, per-year output and recent-venue tempo are not available in the input data, so this axis is unscored; today only the 2026 year labels of the two preprints can be confirmed.

Reproducibility100/100

Reproducibility scores 100/100, with a raw reproMean of 100 — but computed from only one paper in the reproducibility sample (reproPapers: 1). Consistently, both catalogued preprints carry a high reproducibility-readiness rating, reflecting reliance on linear probing, activation probes and causal interventions over language models that others can re-run. Evolve's verdicts call the analyses careful while noting modest effect sizes. The perfect score is real but rests on a very small base.

Research summary

Gouki Minegishi is an early-career machine learning researcher at the University of Tokyo whose catalogued work centers on the internal representations and self-monitoring capacities of large language models. Evolve's catalog holds two 2026 arXiv preprints. The first, 'Few-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs', shows that demonstrations drawn from different worlds improve in-context graph prediction and causally engage internal world representations — careful analysis, though the effects are modest. The second, 'In-Context Neurofeedback', tests whether LLMs can control their internal representations through privileged access; effects were small and inconsistent across four models, leaving prior introspection claims unproven, though the null result is not conclusive. On both papers Minegishi holds a middle, non-corresponding authorship position, and both carry a high reproducibility-readiness rating, supporting a 100/100 reproducibility score computed from a single-paper sample. The radar profile is sharply concentrated: focus is 100/100 with all catalogued output in machine learning (LLM interpretability and in-context learning), while collaboration scores 50/100 across 11 distinct co-authors, with Yutaka Matsuo, Yusuke Iwasawa and Takeshi Kojima recurring most often. Output, impact and activity dimensions are not yet scored because citation and volume data are unavailable for these recent preprints; the profile should be read as a high-conviction but early-stage snapshot.

Collaboration pattern

Collaboration is team-based and Tokyo-centric. The raw data lists 11 distinct co-authors, with Yutaka Matsuo, Yusuke Iwasawa and Takeshi Kojima each appearing twice in the frequent-coauthor list, alongside Kohsei Matsutani, Core Francisco Park, Koshiro Aoki, Ryota Takatsuki, Yusuke Haruki and Daisuke Kawahara. Minegishi sits in a middle, non-corresponding position on both catalogued papers, consistent with contributing experiments and analysis within a larger interpretability group rather than leading solo projects.

Research keywords

in-context learningworld modelsmechanistic interpretabilitylinear probingcausal interventionfew-shot promptingLLM metacognitionneurofeedbackprivileged accessintrospectionactivation probesAI safety

Highlights

Few-shot demonstrations unlock in-context world representations
Evolve's review finds that few-shot demonstrations drawn from different worlds boost in-context graph prediction and causally engage the model's internal world representations, with careful analysis — though the effects are modest. The paper is rated high for reproducibility readiness.
A privileged-access neurofeedback test of LLM self-control
The paper proposes a privileged-access neurofeedback test of whether LLMs can control their own internal representations. Effects were small and inconsistent across four models — prior introspection claims are unproven, though this null result is itself not conclusive. Rated high for reproducibility.
A probing-and-intervention toolkit with high reproducibility
Across both catalogued papers, the same methodological core recurs: linear and activation probes plus causal intervention on language models. Both preprints are rated high for reproducibility readiness, and the reproducibility sample yields a raw reproMean of 100 from one rated paper.

Timeline

2026First catalogued arXiv preprint appears: few-shot demonstrations are shown to elicit in-context world representations in LLMs, rated high for reproducibility readiness.
2026Second preprint proposes the in-context neurofeedback test of LLM privileged access across four models, also rated high for reproducibility readiness.
2026With both papers in LLM interpretability and metacognition, the profile records a single-field focus score of 100/100 and an initial network of 11 distinct co-authors.

Papers in the Evolve catalog

Frequent co-authors

Kohsei Matsutani×1 Core Francisco Park×1 Takeshi Kojima×1 Yusuke Iwasawa×1 Yutaka Matsuo×1 Yusuke Iwasawa×1 Yutaka Matsuo×1 Koshiro Aoki×1 Ryota Takatsuki×1 Yusuke Haruki×1 Daisuke Kawahara×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 scores come from a small sample: the focus score rests on a single field bucket and the reproducibility score on one reproducibility-rated paper (reproPapers: 1).
Output, impact and activity axes are unscored because citation and publication-volume data are not yet available for these 2026 arXiv preprints.
No ORCID or country field was provided; the profile is name-resolved from public bibliographic data and Evolve's reviewed papers.
Coverage is limited to the two Evolve-reviewed papers; other publications by this author may exist outside the catalog.