Cambridge researcher in multilingual NLP and computational psycholinguistics: bilingual LMs and activation steering
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) 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—
Not yet scored: Evolve has no reliable publication-volume data for this researcher. This axis would measure total research output — the number of papers, venues, and preprints over time — once the underlying record is complete.
Impact—
Not yet scored: citation counts, h-index, and downstream influence are unavailable in the current data. This axis would measure how widely the two 2026 preprints and any earlier work are cited and built upon by other researchers.
Focus0/100
The radar computes focus at 0, reflecting an even 0.5/0.5 split between multilingual NLP and computational psycholinguistics across the reviewed work. In practice the two halves are tightly coupled — the Beetle suite uses bilingual language models to predict second-language reading times, while LISECO steers language choice in multilingual models — so the low score indexes breadth across two adjacent subfields rather than scattered unrelated work.
Collaboration39/100
Collaboration scores 39, with 6 distinct co-authors recorded across the reviewed papers. The network is small and stable: repeated collaborators include Roger P. Levy, Paula Buttery, Tyler A. Chang, Catherine Arnett, Suchir Salhan, and Carmen Amo Alonso. Both catalog entries are team papers — a middle-authorship suite release and a first-author methods paper — consistent with small close-knit teams rather than large multi-institution consortia.
Activity—
Not yet scored: no year-by-year publication cadence is available beyond the two 2026 preprints. This axis would measure publishing persistence — the regularity and recency of output across the research career — once a fuller record exists.
Reproducibility78/100
Reproducibility is the strongest scored axis at 78, averaged over 2 reviewed papers. The Beetle preprint rates high: it releases 285 bilingual and 45 monolingual language models with checkpoints. LISECO rates medium — careful evaluation, but a narrower release footprint. Together they indicate a habit of shipping artifacts, tempered by noted limitations such as single-seed training in Beetle.
Research summary
James A. Michaelov is a researcher at the University of Cambridge working at the intersection of multilingual NLP and computational psycholinguistics. Evolve's catalog covers two 2026 arXiv preprints, split evenly across these two fields (0.5/0.5). The first, Beetle (middle authorship), releases a bilingual model suite — 285 bilingual and 45 monolingual language models with checkpoints — and finds that staged curricula best predict second-language reading times, though the review notes that single-seed training and L2-share confounds temper the claims. The second, LISECO (first authorship), frames target-language generation in multilingual models as activation steering cast as optimal control: carefully evaluated, but outperforming a tuned difference-in-means baseline on only 5 of 8 models in a single task. Radar scores show strong reproducibility (78, averaged over 2 papers), driven by substantial artifact releases and careful evaluation; a focus score of 0 reflects an even split across two closely related subfields rather than scattered work; and a modest collaboration score (39) corresponds to six distinct co-authors in a stable recurring set including Roger P. Levy, Paula Buttery, Tyler A. Chang, Catherine Arnett, Suchir Salhan, and Carmen Amo Alonso. Output, impact, and activity lack sufficient data and are not yet scored.
Collaboration pattern
Collaboration scores 39 with 6 distinct co-authors — a small, stable network rather than high-churn teamwork. Frequent collaborators are Roger P. Levy, Paula Buttery, Tyler A. Chang, Catherine Arnett, Suchir Salhan, and Carmen Amo Alonso. The pattern mixes psycholinguistics-oriented modeling work (Beetle, middle authorship) with first-authored control and steering methods (LISECO), suggesting tight-knit lab-style collaboration where the same core group recurs across projects.
Research keywords
Highlights
Beetle: 285 bilingual + 45 monolingual LMs released with checkpoints
Staged curricula best predict L2 reading times
LISECO: language steering as optimal control
Timeline
- 2026First-authored LISECO, framing target-language generation in multilingual models as activation steering with optimal control (arXiv).
- 2026Co-authored (middle position) Beetle, releasing 285 bilingual and 45 monolingual language models with checkpoints for second-language processing; Evolve rates reproducibility readiness high.
- 2026Both reviewed papers land in 2026, split evenly between multilingual NLP and computational psycholinguistics, with reproducibility averaging 78 across the two.
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.
- Radar dimensions are computed from public bibliometric data only; output, impact, and activity axes lack sufficient data and are unscored.
- Evolve's reviewed catalog contains only two arXiv preprints, both from 2026; they may not represent the full publication record.
- The reproducibility score (78) averages only 2 papers, so it may not generalize to the wider body of work.
- Profile is name-resolved from public bibliographic data (ORCID 0000-0003-2913-1103); biographical details beyond the listed affiliation are unverified.