Yutaka Matsuo

The University of Tokyo · Japan

Unclaimed · AI-structured profileOpenAlex

University of Tokyo researcher bridging LLM in-context learning and vision-language-action robot manipulation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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—

Output is not scored yet: the snapshot provides per-year counts (1 work in 2025, 46 in 2026) but no validated full works list or venue metadata, so total publication volume and venue quality cannot be assessed on this axis.

Impact—

Impact is not scored yet: citation coverage is far too thin — the record shows only an h-index of 1 and an i10-index of 0 — which cannot represent real influence, so this axis stays unscored until citation data accumulates.

Focus0/100

Focus scores 0 because output splits exactly evenly between two fields: robot learning (vision-language-action models, manipulation) and machine learning (LLM interpretability and in-context learning), each at 0.5 of the field distribution. Topic counts confirm the breadth — Topic Modeling (28), Natural Language Processing Techniques (18), Adversarial Robustness (13), Multimodal Applications (11), Reinforcement Learning in Robotics (8) — pointing to a deliberate two-track program rather than a narrow specialty.

Collaboration48/100

Collaboration scores 48 with 10 distinct coauthors in the snapshot. Yusuke Iwasawa is the only name appearing twice in the frequent-coauthor list, marking the densest recurring tie; other recurring names include Takeshi Kojima, Tatsuya Matsushima, Maxime Alvarez, Renzo Caballero, Core Francisco Park, Kohsei Matsutani and Gouki Minegishi. On both reviewed 2026 papers Matsuo sits in the last author position, a pattern typical of a senior author coordinating a multi-member lab group.

Activity100/100

Activity scores 100, the maximum. countsByYear records 1 work in 2025 and 46 in 2026, and the recent three-year window totals 47 works — an exceptional publication rate concentrated almost entirely in 2026, indicating a steep output ramp and an intensely productive current phase spanning both research tracks.

Reproducibility78/100

Reproducibility scores 78, the mean across the 2 Evolve-reviewed papers (reproPapers 2). The LLM interpretability study carries high reproducibility readiness thanks to linear probing and causal-intervention protocols, while the latent action model study is rated medium, citing single-run statistics. A 78 average on controlled, protocol-driven empirical work signals solid methodological transparency.

Research summary

Yutaka Matsuo is a researcher at The University of Tokyo whose OpenAlex-resolved record splits almost evenly between two fields: robot learning (vision-language-action models, manipulation) and machine learning (LLM interpretability and in-context learning), each at 0.5 of the field distribution. The snapshot is unusually output-heavy — countsByYear lists 1 work in 2025 and 46 in 2026, with 47 works in the recent three-year window, earning a maximal activity score of 100 — while citation-derived indicators remain extremely thin (h-index 1, i10-index 0), so impact cannot be scored. Topic mass concentrates in Topic Modeling (28), Natural Language Processing Techniques (18), Adversarial Robustness (13), Multimodal ML Applications (11) and Reinforcement Learning in Robotics (8). Evolve has reviewed two 2026 arXiv papers, both with Matsuo in the last author position: a mechanistic-interpretability study showing that few-shot demonstrations from different worlds elicit in-context world representations in LLMs (high reproducibility readiness), and a latent-action-model study lifting cross-embodiment transfer on RoboTwin 2.0 from 24% to 61.3% via action-similarity supervision (medium). The two assessed papers average a reproducibility score of 78. Collaboration sits mid-range (score 48, 10 distinct coauthors), with Yusuke Iwasawa the only repeated coauthor name in the list.

Collaboration pattern

The snapshot records 10 distinct coauthors and a mid-range collaboration score of 48. Yusuke Iwasawa is the only name appearing twice in the frequent-coauthor list, marking the densest recurring tie; other recurring names include Takeshi Kojima, Tatsuya Matsushima, Maxime Alvarez, Renzo Caballero, Core Francisco Park, Kohsei Matsutani and Gouki Minegishi. On both reviewed 2026 papers Matsuo occupies the last author position, a pattern consistent with a senior author coordinating a multi-member lab group.

Research keywords

in-context learningworld modelsmechanistic interpretabilityvision-language-action modelslatent action modelscross-embodiment transferrobot manipulationfew-shot promptinglinear probingcausal intervention

Highlights

Few-shot demonstrations elicit in-context world representations in LLMs
Evolve's review finds that demonstrations drawn from different worlds boost in-context graph prediction and causally engage world representations, with linear probing and causal-intervention evidence; the analysis is careful, though effects are modest, and reproducibility readiness is rated high.
Action-similarity supervision lifts latent action model transfer on RoboTwin 2.0
The study raises cross-embodiment transfer of latent action models from 24% to 61.3% on RoboTwin 2.0 using action-similarity supervision with relational knowledge distillation; Evolve flags single-run statistics and a competitive InfoNCE baseline, rating reproducibility readiness medium.

Timeline

2025The record opens with 1 work in 2025, the starting point of a steep publication ramp captured in countsByYear.
2026Output surges to 46 works in 2026 (47 over the recent three-year window), earning a maximal activity score of 100.
2026Last-author arXiv study shows few-shot demonstrations from different worlds elicit in-context world representations in LLMs, with high reproducibility readiness.
2026Last-author arXiv study lifts latent action model cross-embodiment transfer on RoboTwin 2.0 from 24% to 61.3%, rated medium for reproducibility.

Papers in the Evolve catalog

Frequent co-authors

Yusuke Iwasawa×2 Maxime Alvarez×1 Renzo Caballero×1 Tatsuya Matsushima×1 Kohsei Matsutani×1 Gouki Minegishi×1 Core Francisco Park×1 Takeshi Kojima×1 Yusuke Iwasawa×1 Yutaka Matsuo×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.

Identity is name-resolved from public OpenAlex bibliographic data only; no ORCID link was available, and the name is a common Japanese one, so homonym mixing is possible.
Citation indicators in the snapshot (h-index 1, i10-index 0) are implausibly thin beside the 47 recent works reported, so the output and impact axes are left unscored rather than estimated.
The reproducibility score (78) averages only the 2 Evolve-reviewed papers; the remaining works among the 47 recent outputs are unreviewed.
The frequent-coauthor list includes 'Yutaka Matsuo' himself, a de-duplication artifact, so co-author tallies may shift after cleanup.