Computational social scientist auditing LLMs as social and political actors — herding, bias, alignment reactivity.
3
papers here
3
reviewed by Evolve
3
first-author
3
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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: the bibliometric record is too sparse to compute a reliable output measure. This axis would aggregate publication volume, venue tier, and growth trajectory; the current snapshot shows only an h-index of 1 and i10-index of 0.
Impact—
Not yet scored: citation-based impact cannot be estimated from a record this thin (i10-index 0, h-index 1). This axis would measure citations, scholarly uptake, and real-world influence; it should be revisited once the 2026 preprints accrue citations.
Focus0/100
The focus score of 0 stems from an even three-way split of the field distribution — 'ai safety / computational social science', 'computational social science', and 'computational social science / nlp auditing' each hold about one third — while topics spread across five categories led by Ethics and Social Impacts of AI (9) and International Relations and Foreign Policy (4). The reviewed papers themselves read as one coherent LLM-auditing agenda, so the low score likely reflects taxonomy granularity more than genuine diffusion.
Collaboration14/100
The collaboration score of 14 reflects a distinctly solo pattern: across 12 works in the last three years, OpenAlex records only 1 distinct coauthor, and the frequent-coauthor list contains no other researcher. All three reviewed preprints position Chupilkin as first and corresponding author. Some of this may be name-disambiguation, but on current evidence this reads as independent scholarship that has not yet run through co-authored teams.
Activity75/100
The activity score of 75 is driven by a sharp recent acceleration: the yearly record moves from 1 work in 2019 and 1 in 2021 to 2 in 2025 and 10 in 2026, with 12 works in the last three years — nearly all visible output lands in the most recent year. The pattern points to an intensive current phase, consistent with the 2026 arXiv preprint series.
Reproducibility55/100
Reproducibility scores 55, a mean across the 3 papers Evolve has assessed (reproMean 55), each rated 'medium' reproducibility readiness. The Evolve verdicts flag concrete limits behind the mid-range score: a single-model community sample in the LLM-herding study, descriptive rather than causal evidence in the 112-language audit, and a single-cue confound with floor effects in the alignment experiment.
Research summary
Maxim Chupilkin, listed at the European Bank for Reconstruction and Development (UK), works at the intersection of computational social science, AI safety, and NLP auditing, treating large language models as objects of social-scientific measurement rather than only as tools. The OpenAlex record shows an early-stage but rapidly accelerating profile: an h-index of 1 and i10-index of 0, yet 12 works in the last three years, 10 of them in 2026 — the most active year on record. All three papers Evolve has reviewed are 2026 arXiv preprints with Chupilkin as first and corresponding author. One transplants the classic 'Music Lab' design into LLM-agent populations, finding that social counts cut selections by 17.2% while a random head start adds 45.55 percentage points. A second audits large language models across 112 languages, showing that answers on the Russia–Ukraine war shift with prompt language and track global politics. The third uses a conjoint experiment to show a single alignment cue lowers LLM willingness to go to war by 13.4 points and reshapes decision rules. Topic data — Ethics and Social Impacts of AI (9 works), International Relations and Foreign Policy (4) — points to one coherent agenda: how LLM behavior bends under social, geopolitical, and evaluative pressure.
Collaboration pattern
On current OpenAlex evidence Chupilkin works independently: 12 recent works list just 1 distinct coauthor, and the frequent-coauthor list contains only Chupilkin himself, giving a collaboration score of 14. All three reviewed preprints are first- and corresponding-authored. The pattern may partly reflect name disambiguation in bibliographic databases rather than genuinely isolated work; co-authorship metrics should be read as a lower bound until records are consolidated.
Research keywords
Highlights
Music Lab goes machinic: social influence reshapes LLM agents' attention choices
One war, 112 languages: geopolitical divisions across LLMs
Alignment tests change the verdict: LLMs judge war differently when observed
Timeline
- 2019Earliest work appears in the OpenAlex record (1 work).
- 2021One further work recorded — output remains sparse in this period.
- 2025Two works recorded, marking the start of a ramp-up.
- 2026Most productive year on record: 10 works, including all three reviewed arXiv preprints, each first- and corresponding-authored.
Papers in the Evolve catalog
- first authorcorresponding
Social Influence and the Allocation of Scientific Attention in AI Populations
- first authorcorresponding
Geopolitical Divisions Across Languages in Large Language Models
- first authorcorresponding
Language models judge war differently when tested for alignment
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 indicators come from OpenAlex only; no Scopus, Web of Science, or Google Scholar data was used.
- Output and impact axes are unscored because citation coverage is sparse (h-index 1, i10-index 0), and all reviewed papers are 2026 arXiv preprints with little time to accumulate citations.
- The record shows only 1 distinct coauthor and a frequent-coauthor list containing the researcher alone; name disambiguation may undercount real collaborators.
- Reproducibility and quality judgments quote Evolve's review verdicts on 3 papers, not independent peer reviews; all three carry 'medium' reproducibility readiness.