Tokyo researcher bridging robot learning and LLM interpretability: VLA transfer and in-context world representations
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
16
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—
No output score is computed yet: this axis would measure total and per-year publication volume with venue- and author-role weighting. Raw counts exist (91 works over the last three years), but the normalized scoring data is not available.
Impact—
Impact scoring is pending: this axis measures citations, field-normalized citation percentiles, and citing works. Scholar counts (h-index 16, i10-index 26) are recorded, but per-paper citation data required for the score is missing.
Focus0/100
The focus score of 0 reflects an evenly split agenda: the field distribution is exactly 50% robot learning (vision-language-action models, manipulation) and 50% machine learning (LLM interpretability and in-context learning). Topic counts confirm breadth over concentration — topic modeling (57), domain adaptation and few-shot learning (41), NLP techniques (39), reinforcement learning in robotics (32), and adversarial robustness (29) — spanning at least five clusters.
Collaboration48/100
A collaboration score of 48 sits mid-range. The record lists 10 distinct coauthors, with a recurring Tokyo cluster: Yutaka Matsuo (appearing twice in the frequency list), Takeshi Kojima, Tatsuya Matsushima, plus Maxime Alvarez, Renzo Caballero, Kohsei Matsutani, Gouki Minegishi, and Core Francisco Park. Both Evolve-reviewed 2026 papers place this researcher in a middle author position without corresponding role, consistent with contributing within a team.
Activity100/100
Activity scores 100, the maximum. The record shows 91 works in the last three years, and by-year counts rise steeply: 8 (2020), 15 (2021), 15 (2022), 14 (2023), 17 (2024), 34 (2025), 40 (2026) — output has more than doubled since 2024. Even the earliest recorded years (1 work each in 2011 and 2012) show a continuous, now sharply accelerating publishing cadence.
Reproducibility78/100
Reproducibility averages 78 across the two papers Evolve has assessed. The 2026 paper on few-shot demonstrations and in-context world representations is rated high reproducibility readiness, while the latent-action-model cross-embodiment paper is rated medium, with the review noting single-run statistics and a rival InfoNCE baseline. A 78 mean on a two-paper base is promising but not yet a stable estimate.
Research summary
Yusuke Iwasawa is a researcher at The University of Tokyo whose recorded work splits evenly — a 50/50 field distribution — between robot learning (vision-language-action models, manipulation) and machine learning centered on LLM interpretability and in-context learning. The OpenAlex-based record shows 91 works in the last three years, with annual counts climbing from 17 in 2024 to 34 in 2025 and 40 in 2026, producing an activity score of 100. Recurring topics include topic modeling (57), domain adaptation and few-shot learning (41), natural language processing techniques (39), reinforcement learning in robotics (32), and adversarial robustness (29). Scholar counts list an h-index of 16 and i10-index of 26, though the output and impact dimension scores remain uncomputed. Two Evolve-reviewed 2026 papers illustrate the dual agenda: one shows few-shot demonstrations eliciting in-context world representations in LLMs via probing and causal interventions (high reproducibility readiness); the other uses action-similarity supervision to lift latent-action-model cross-embodiment transfer on RoboTwin 2.0 from 24% to 61.3% (medium readiness). A reproducibility score of 78, averaged over two assessed papers, and a mid-range collaboration score of 48 with 10 distinct coauthors round out the profile. The record's country field ('US') conflicts with the listed institution and is not asserted here.
Collaboration pattern
The record lists 10 distinct coauthors and a mid-range collaboration score of 48. The recurring core is a Tokyo cluster — Yutaka Matsuo (appearing twice in the frequency list), Takeshi Kojima, and Tatsuya Matsushima — alongside Maxime Alvarez, Renzo Caballero, Kohsei Matsutani, Gouki Minegishi, and Core Francisco Park. On both Evolve-reviewed 2026 papers this researcher holds a middle author position without corresponding role, indicating contribution within a larger team; leadership patterns cannot be assessed from available data.
Research keywords
Highlights
Few-shot prompts elicit in-context world representations in LLMs
Action-similarity supervision lifts latent action model transfer
A two-track agenda linking embodied AI and LLM science
The two reviewed 2026 papers sit at opposite ends of a 50/50 field split between robot learning and LLM interpretability, yet share a methodological core: probing and shaping internal representations, whether for world models in language models or action representations in vision-language-action systems.
Timeline
- 2011First year in the bibliometric record, with 1 recorded work; annual counts stay low (1–3) through 2017.
- 2018Recorded works jump to 15, an early inflection from the preceding single-digit annual counts.
- 2021Sustained high output begins: 15 recorded works in 2021 and 15 again in 2022, with 14 in 2023.
- 2025The annual count reaches 34, more than double 2024's 17 recorded works.
- 2026Recorded works reach 40, and two Evolve-reviewed papers appear: few-shot-elicited in-context world representations in LLMs, and action-similarity supervision for latent action models.
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 figures (works, topics, coauthors, identity) come from OpenAlex records; the profile is name-resolved from public bibliographic data.
- Output and impact dimension scores are missing; only raw scholar counts (h-index 16, i10-index 26) are available, without per-paper citation distributions.
- The reproducibility score of 78 averages only two Evolve-assessed papers (one high, one medium readiness), a thin base.
- The record's country field says 'US' while the listed institution is The University of Tokyo; the conflict is unresolved and no country is asserted.