CNRS natural language processing researcher: factual recall in LLMs and rigorous evaluation of explainability methods
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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: Evolve holds no publication-count or venue-volume data, so the output axis is empty. It would measure how much publishable work he produces over time — total papers, annual volume, and authorship roles.
Impact—
Not yet scored: citation and influence metrics are absent from the snapshot, so impact cannot be estimated. This axis would measure how far the work propagates — citations, downstream adoption, and uptake of the findings by other groups.
Focus0/100
The focus score is 0, derived from a field distribution that splits evenly: 0.5 natural language processing and 0.5 natural language processing (explainable AI evaluation). In practice the two labels describe adjacent territory — both catalog papers probe how large language models behave under evaluation, one on factual recall and memorization, one on explainability assessment — so the split reflects label granularity as much as genuine breadth, but the raw data records no single dominant field.
Collaboration46/100
Collaboration scores 46, a moderate value backed by 9 distinct coauthors in the snapshot. The frequent-coauthor list pairs Philippe Muller with a cluster of researchers around the explainability and evaluation work — Guilhem Fouilhé, Antonin Poché, Fanny Jourdan, Nils Feldhus, Qianli Wang, Jing Yang, Simon Ostermann, and Vera Schmitt — a pattern of recurring, relatively small teams rather than a broad rotating network, which fits the mid-range score.
Activity—
Not yet scored: no publication-frequency or recency time series exists, so the axis stays unscored. It would measure how consistently he publishes in recent years — whether the 2026 arXiv papers mark a sustained phase or a narrow window.
Reproducibility100/100
Reproducibility is the strongest measured axis: a mean of 100 over 2 repro-scored papers, both rated 'high' in reproducibility readiness. The two 2026 studies are consistent with this — the factual-recall replication spans 31 models for others to re-run, and the simulatability critique pairs its negative findings with actionable fixes — indicating procedures designed so independent teams can verify the results.
Research summary
Nicholas Asher is a researcher at the Centre National de la Recherche Scientifique (CNRS) whose cataloged recent work sits in natural language processing, splitting evenly in the field distribution (0.5 general NLP and 0.5 NLP for explainable AI evaluation). The Evolve catalog holds two 2026 arXiv papers, in both of which he is a middle author without the corresponding role. The first, 'Memory vs. Context?', replicates and stress-tests Yu et al. (2023) across 31 models, finding that size and frequency trends mostly hold while phrasing alone can swing memorization by up to 95 percentage points. The second, 'Limitations of Automated Simulatability', is a persuasive self-critique showing that LLM simulators can bypass concept-based explanations by leaning on task priors and leaking anonymized labels, with mostly correlational evidence on models up to 15B parameters. Both papers carry a 'high' reproducibility readiness rating, producing a reproducibility mean of 100 over 2 scored papers. The collaboration snapshot lists 9 distinct coauthors — including Philippe Muller, Nils Feldhus, and Qianli Wang — with a moderate collaboration score of 46. Output, impact, and activity axes lack sufficient data and remain unscored, so this profile is an early, evidence-bounded view built from the reviewed papers alone.
Collaboration pattern
The snapshot lists 9 distinct coauthors and a moderate collaboration score of 46. Philippe Muller appears alongside a cluster of frequent collaborators on the explainability work — Guilhem Fouilhé, Antonin Poché, Fanny Jourdan, Nils Feldhus, Qianli Wang, Jing Yang, Simon Ostermann, and Vera Schmitt. In both catalog papers Asher takes a middle authorship without the corresponding role, consistent with contributing to a small, recurring team rather than solo or led projects.
Research keywords
Highlights
Cross-model replication of factual recall trends
Showing LLM simulators can bypass explanations
Two high-reproducibility evaluation studies in 2026
Both catalog papers earn the top reproducibility rating, giving a reproducibility mean of 100 over 2 scored papers — a small but consistent record of evaluation work designed for independent verification.
Timeline
- 2026Middle-author arXiv paper on factual recall: a replication across 31 models finds phrasing alone shifts memorization by up to 95 percentage points.
- 2026Middle-author arXiv self-critique of automated simulatability, showing LLM simulators can bypass explanations via task priors and label leakage.
- 2026Both papers enter the Evolve catalog rated high for reproducibility, lifting the reproducibility mean to 100 over 2 scored papers.
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 scores are computed from OpenAlex-derived data only; the output, impact, and activity axes lack sufficient data and are left unscored.
- Profile content draws on only 2 reviewed papers, both 2026 arXiv preprints with middle authorship — an early, evidence-bounded view.
- 'Nicholas Asher' is a common-style name; the profile is name-resolved from public bibliographic data and may not cover all work by researchers with the same name.
- Country-level affiliation information is missing in the current snapshot.