IRIT Toulouse NLP researcher studying factual recall, memorization, and explainability evaluation in LLMs
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—
Publication-volume data is not available in this profile's bibliometric snapshot, so output is left unscored. This axis would measure the number and cadence of papers over time, venue tiers, and the share of first- or corresponding-author outputs.
Impact—
No citation-side data is present for this researcher yet, so impact is unscored. The axis would measure citation counts and aggregates, and the eventual uptake of the two 2026 arXiv preprints once citation histories accumulate.
Focus0/100
The focus score is 0, driven by an even field distribution: the raw split is exactly 0.5 natural language processing and 0.5 natural language processing (explainable AI evaluation). In practice this looks narrower than the score suggests — both 2026 catalog papers sit inside NLP, one on factual recall and memorization using logit lens, the other on explainability evaluation — but the exact two-way tie between subfields keeps the computed focus score at the floor.
Collaboration46/100
The collaboration dimension scores 46, computed from 9 distinct co-authors in the raw snapshot. The frequent list pairs Nicholas Asher, Guilhem Fouilhé, Antonin Poché, and Fanny Jourdan with Nils Feldhus, Qianli Wang, Jing Yang, Simon Ostermann, and Vera Schmitt, a nine-name network behind a small body of co-authored work. On the two catalog papers Muller sits in middle and last author positions, neither as corresponding author, consistent with small-team co-authorship.
Activity—
Year-level publication activity cannot be computed from the available data, so this axis is unscored. It would measure recent output cadence, gaps between publications, and whether research activity is rising or tapering across recent years.
Reproducibility100/100
This is the strongest axis: the reproducibility score is 100, with reproMean 100 computed over the 2 papers in the catalog. Both entries are flagged high reproReadiness by Evolve. The factual-recall paper is itself an explicit replication and stress-test of Yu et al. (2023) across 31 models — a design that directly supports re-execution — while the simulatability critique, though mostly correlational in its evidence, proposes concrete, actionable fixes to its evaluation protocol.
Research summary
Philippe Muller is a natural language processing researcher at the Institut de Recherche en Informatique de Toulouse (IRIT). The work Evolve has reviewed concentrates on two adjacent questions about large language models: what a model actually knows parametrically versus what it picks up from context, and whether the tools we use to evaluate explanations actually measure the explanations themselves. The 2026 catalog holds two arXiv papers. On 'Memory vs. Context?' he is the last author: a replication and stress-test of Yu et al. (2023) across 31 models showing that model-size and frequency trends in factual recall mostly hold, while phrasing alone swings memorization by up to 95 percentage points. On 'Limitations of Automated Simulatability' he is a middle author on a self-critique showing that LLM simulators can bypass explanations by leaning on task priors and leaking anonymized labels. Both papers carry Evolve's highest reproducibility readiness, and the snapshot's reproducibility dimension scores 100 (reproMean 100 over 2 papers). Field distribution splits evenly at 0.5/0.5 between natural language processing and explainable-AI evaluation, with 9 distinct co-authors recorded; the two-subfield tie drives the focus score of 0.
Collaboration pattern
Collaboration is small-team and repeat-partner based: 9 distinct co-authors appear in the raw snapshot, giving a collaboration score of 46. Frequent partners include Nicholas Asher, Guilhem Fouilhé, Antonin Poché, and Fanny Jourdan, alongside Nils Feldhus, Qianli Wang, Jing Yang, Simon Ostermann, and Vera Schmitt. Across the two catalog papers Muller takes middle and last author positions rather than the corresponding slot, indicating shared authorship rather than sole correspondence.
Research keywords
Highlights
31-model replication of factual recall findings
Showing LLM simulators can bypass explanations
High reproducibility across the catalog
Both reviewed papers rate high on reproducibility readiness, giving reproMean 100 over the 2 catalog papers and a reproducibility dimension score of 100.
Timeline
- 2026Last author on 'Memory vs. Context? Influential Factors of Factual Recall in Language Models', a 31-model replication and stress-test of factual recall (arXiv).
- 2026Middle author on 'Limitations of Automated Simulatability: LLM Simulators Can Bypass Explanations', documenting shortcuts in explainability evaluation (arXiv).
- 2026Both papers enter the Evolve catalog with high reproducibility readiness; the snapshot's reproMean is 100 over these 2 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 derive from OpenAlex-based data; the output, impact, and activity dimensions lack sufficient underlying data and are left unscored.
- The profile rests on 2 arXiv preprints, both from 2026; no peer-reviewed venue, citation, or funding information is included in the input data.
- 'Philippe Muller' is a common name; this profile is name-resolved from public bibliographic data anchored to ORCID 0000-0002-6765-4020 and the IRIT affiliation.
- Co-author affiliations and the researcher's country are not recorded in the input data, so collaboration partners are identified by name only.