Alexandre Quemy

—

Unclaimed · AI-structured profileORCID OpenAlex

Language-model interpretability researcher charting transformer computation via geometry and optimal transport

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—

The output axis is unscored: the snapshot provides an h-index of 6 and i10-index of 3 but no complete works list, so total volume and mix of outputs cannot be established from the data available.

Impact—

The impact axis is unscored: per-paper citation counts are absent from the snapshot, so influence on the literature cannot be measured; the index values of 6 and 3 hint at modest uptake but do not support a score.

Focus100/100

Focus scores a maximum 100: the reviewed output sits in a single field, machine learning — language model interpretability, and all four catalog papers attack transformer internals from geometric angles. The broader topic record — topic modeling (7), XAI (7), AI in law (6) — stays within one thematic band, signaling a deliberately narrow agenda.

Collaboration0/100

The snapshot records zero distinct co-authors and an empty frequent-collaborator list, yet all four cataloged papers list him as first and corresponding author — a pattern consistent with solo-led work. Whether this reflects genuinely single-author research or unpopulated co-author linkage in the OpenAlex record cannot be determined from the data; the zero score reflects the missing links as much as the writing pattern.

Activity57/100

Activity scores 57 on a mixed cadence: yearly counts stay low through 2013–2021 (one to three works) before jumping to seven in 2023 and six in 2026, with six works in the last three years. The recent cluster is entirely interpretability preprints, first- and corresponding-authored, indicating a strong current push after quieter earlier years.

Reproducibility55/100

Reproducibility scores 55 from a thin base: only one paper carries a computed score (reproMean 55), and both slugged catalog entries — the task-weighted chart paper and the optimal transport study — carry medium repro-readiness. Evolve's verdict on the latter calls it sound but thin, using two small models, which fits the middling score.

Research summary

Alexandre Quemy (ORCID 0000-0002-5865-6403; profile name-resolved from OpenAlex bibliographic data; institution and country unrecorded) is a machine-learning researcher whose current work concentrates on the interpretability of language models, examined through geometric and measure-transport tools. Four 2026 arXiv preprints in the Evolve catalog — all first- and corresponding-authored — form a tight series: task-weighted charts of a language model's computation built on intrinsic dimension and the Fisher information metric; an optimal-transport analysis of transformer depth on Pythia-160m/410m; a nonlinear characterization of the depth flow of token representations; and a study of how a hub of short rows inflates intrinsic-dimension estimates of token embeddings. Evolve's verdict on the first finds the 70–90% residual-width requirement for next-token functions persuasive in tooling while flagging dependence on tolerance choices; the second is called sound but thin, covering two small models. The bibliometric snapshot backs a focused rather than prolific record: h-index 6, i10-index 3, yearly counts of one to three works through 2013–2021 rising to seven in 2023 and six in 2026. Earlier topic tags — topic modeling, explainable AI, AI in law, text classification — sit alongside the present interpretability line, suggesting a coherent applied-AI trajectory now narrowed to transformer internals. Output and impact dimensions remain unscored for lack of complete works and citation data.

Collaboration pattern

The snapshot lists no frequent co-authors and counts zero distinct collaborators, while all four catalog papers are first- and corresponding-authored arXiv preprints — most plausibly solo work. A caveat applies: co-author linkage comes from the OpenAlex identity record and may be incomplete, so the absence of listed collaborators should be read as 'none recorded' rather than a confirmed solo career.

Research keywords

intrinsic dimensionFisher information metricresidual streamoptimal transporttransformer interpretabilityBrenier mapWasserstein distancemodel compressionknowledge distillationtopic modelingtext classificationexplainable AI

Highlights

Task-weighted charts: the next-token function needs 70–90% of residual width
Evolve's review of this first-authored 2026 preprint finds the claim that the next-token function requires 70–90% of the residual stream width — none of it captured by variance — backed by persuasive experimental tooling, while cautioning that the core numbers depend on tolerance choices. Builds on intrinsic dimension and the Fisher information metric, with links to compression and distillation.
Optimal transport emerges in the final layers of trained Pythia models
A first- and corresponding-authored 2026 study measures optimal transport across transformer depth: trained Pythia-160m/410m move final-layer tokens along the OT map, first layers do not, and the middle remains unresolved. Evolve calls it sound but thin — two small models — with medium repro-readiness.
The depth flow of token representations is nonlinear
A 2026 first-authored arXiv preprint arguing that the depth-wise flow of token representations is nonlinear and does not descend its own density, extending the geometric program of the catalog's other papers. It has no Evolve verdict or record slug yet.
A caution for intrinsic-dimension estimation on token embeddings
Another 2026 first-authored preprint shows that a hub of short rows inflates intrinsic-dimension estimates of token embeddings — a methodological warning directly relevant to the intrinsic-dimension tooling used in the task-weighted charts paper. No Evolve verdict yet.

Timeline

2013First recorded work in the bibliometric snapshot (1 publication).
2015Two recorded publications in the snapshot's counts by year.
2018Three recorded works, starting a steady 2018–2021 run of two to three per year.
2023Peak year in the snapshot: seven recorded works.
2026Six works recorded, including four first- and corresponding-authored arXiv preprints on language-model interpretability in the Evolve catalog.

Papers in the Evolve catalog

2026-09-17first authorcorresponding
Function Lives Where Variance Doesn't: Task-Weighted Charts of a Language Model's Computation
2026-09-01first authorcorresponding
Measuring Optimal Transport in Transformer Depth

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 (h-index 6, i10-index 3, counts by year, topics) come from an OpenAlex identity record; institution and country are not recorded in the input.
Output and impact dimensions are unscored because complete works and per-paper citation data are absent from the snapshot.
Reproducibility rests on a single assessed paper (reproMean 55); two further 2026 preprints lack Evolve verdicts and record slugs.
Zero distinct co-authors may reflect incomplete co-author linkage in the source record rather than genuinely solo authorship.