Machine learning researcher at Lyon 1 Université spanning PAC-Bayesian learning theory and bias-aware LLM compression.
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—
No publication-volume data is available in the current snapshot, so this axis is not scored. It would measure total and annual paper counts across venues, giving a view of overall research productivity and how it has developed over the researcher's career.
Impact—
Citation-based indicators such as citation counts and h-index are absent from the provided data, so this axis cannot be scored. It would measure how widely the reviewed work is cited and taken up by other researchers and by adjacent fields.
Focus0/100
The field distribution is exactly 50% natural language processing (LLM compression and fairness) and 50% machine learning theory (PAC-Bayesian learning), which yields a focus score of 0 — the portfolio is evenly split rather than concentrated. This reads as breadth: one line proves PAC-Bayesian guarantees for majority votes, the other engineers bias-aware pruning for large language models.
Collaboration46/100
The snapshot records 9 distinct coauthors, and the frequent-coauthor list matches that count: Pascal Germain, Amaury Habrard, Emilie Morvant, Paul Viallard, Irina Proskurina, Antoine Gourru, Julien Velcin, Julien Bastian, and Benjamin Leblanc. The collaboration score of 46 sits in the moderate range, and both cataloged 2026 papers list Metzler in a middle, non-corresponding author position.
Activity—
Year-by-year publication records are not available in the snapshot, so this axis remains unscored. It would measure recent publishing cadence — the volume and continuity of outputs over recent years — rather than the two cataloged 2026 entries alone.
Reproducibility55/100
Across the 2 papers Evolve has reviewed, the mean reproducibility-readiness score is 55, with both rated medium. The PAC-Bayesian disintegration paper carries detailed theory whose practical reuse needs care, and Debias-SparseGPT reports real UnQover gains at 1:4 and 50% sparsity, though its 2:4 StereoSet-only results invite replication.
Research summary
Guillaume Metzler is a machine learning researcher affiliated with Lyon 1 Université whose profile, name-resolved from public bibliographic data, spans two equally weighted research lines: natural language processing focused on LLM compression and fairness, and machine learning theory centered on PAC-Bayesian learning. The Evolve catalog holds two 2026 arXiv papers with Metzler in a middle, non-corresponding author position. The first, "On the disintegration of the stochastic majority vote," derandomizes stochastic majority votes through disintegrated PAC-Bayesian bounds and derives a self-bounding algorithm; Evolve judges the theory sound but incremental, noting that the multiclass bound covers a mixture of models rather than one model. The second, "Debias-SparseGPT," adds a bias-aware Hessian to SparseGPT-style pruning; UnQover gains are real at 1:4 and 50% sparsity, yet 2:4 results measured only on StereoSet contradict the paper's "consistently" claim. The bibliometric snapshot records 9 distinct coauthors — among them Pascal Germain, Amaury Habrard, Emilie Morvant, Paul Viallard, Irina Proskurina, Antoine Gourru, Julien Velcin, Julien Bastian, and Benjamin Leblanc — and a moderate collaboration score of 46. Reproducibility readiness averages 55 across the two reviewed papers, both rated medium. Output, impact, and activity axes lack sufficient data, so this early-stage profile should be read as a first, carefully scoped picture rather than a complete record.
Collaboration pattern
The snapshot records 9 distinct coauthors, matching the frequent-coauthor list: Pascal Germain, Amaury Habrard, Emilie Morvant, Paul Viallard, Irina Proskurina, Antoine Gourru, Julien Velcin, Julien Bastian, and Benjamin Leblanc. The collaboration score of 46 is moderate, and both 2026 cataloged papers place Metzler in a middle, non-corresponding author position — a contributor pattern consistent with team-based work at the Lyon 1 Université affiliation.
Research keywords
Highlights
Derandomizing the stochastic majority vote with disintegrated PAC-Bayes
A bias-aware Hessian for SparseGPT pruning
One portfolio, two registers: proofs and pruning
The cataloged output splits evenly (50/50) between PAC-Bayesian theory and LLM compression-and-fairness work, a pattern reflected both in the 50/50 field distribution and in the middle-author roles on both 2026 preprints.
Timeline
- 2026arXiv preprint on disintegrating the stochastic majority vote, from PAC-Bayesian bounds to a self-bounding algorithm (middle author).
- 2026arXiv preprint Debias-SparseGPT: bias-aware pruning for large language models (middle author).
- 2026Both cataloged outputs land in the same year, splitting evenly between PAC-Bayesian theory and LLM compression-and-fairness research.
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 dimensions come from a limited OpenAlex-based snapshot; output, impact, and activity axes lack sufficient data and are not scored.
- No ORCID or country data was available; the profile is name-resolved from public bibliographic records only.
- The evidence base covers just two 2026 arXiv preprints with middle, non-corresponding authorship; peer-reviewed venue versions are not represented.
- The reproducibility mean (55) rests on only two reviewed papers, and all timeline entries fall in the single year 2026.