LMU München researcher building exact, reproducible training-stack methods for Mixture-of-Experts LLMs
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
4
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 citation record is too thin for a stable output measure — h-index 4, i10-index 2, and sixteen works since 2022. This axis would aggregate total publications, per-year output and venue spread once the OpenAlex record matures.
Impact—
Not yet scored: citation-based influence cannot be reliably computed from a record this early. The axis would track citations, h-index trajectory and uptake of the MoE training methods as they accumulate; current evidence is too sparse to estimate.
Focus0/100
Scores 0/100 on concentration, and the raw distributions explain why. The field assignment splits evenly — 0.5 machine learning (large language model training systems) and 0.5 machine learning (large language model training) — while topic counts spread across advanced graph neural networks (10), graph theory and algorithms (6), graph theory and applications (5), topic modeling (4) and advanced neural network applications (3). The profile spans a graph-methods base and LLM training systems rather than one narrow line.
Collaboration44/100
Scores 44/100 with 8 distinct coauthors in the recent window, indicating a small, recurring team rather than a wide network. The frequent-coauthor list is dominated by repeat partners — Pit Neitemeier and Philipp Scholl each appear twice, alongside Jiaze Li, Alessio Serra, Jonas Knupp and Sascha Wirges. Both 2026 catalog papers are multi-author efforts in which Maskey holds first or last position, consistent with a stable core group working on LLM training systems.
Activity75/100
Scores 75/100, the strongest axis in the radar. Counts by year run 3 (2022), 1 (2023), 3 (2024), 3 (2025) and 6 (2026) — sixteen works in total, twelve in the last three years. The 2026 count is the highest on record and includes both Evolve-reviewed MoE papers, so the uptick reflects active, ongoing output rather than a legacy backlog.
Reproducibility38/100
Scores 38/100, derived from a reproducibility mean of 38 over 2 assessed papers. Evolve's readiness ratings split between medium for the quantile-balancing MoE paper and low for the checkpoint-quality study, and both verdicts flag thin evidence bases — one seed and one scale, and a single-seed 30B MoE pipeline. Methods are precisely specified, but reported results are not yet independently robust.
Research summary
Sohir Maskey is a machine learning researcher at Ludwig-Maximilians-Universität München whose recent work concentrates on the systems side of large language model training, especially Mixture-of-Experts (MoE) architectures. The OpenAlex record shows sixteen works between 2022 and 2026, twelve in the last three years, with output accelerating to six works in 2026 — the highest annual count on record. Two 2026 papers reviewed by Evolve attack MoE training from complementary angles: as last author on "Exact Quantile Balancing and Load-Error Injection for Mixture-of-Experts", Maskey contributes to a routing method delivering exact global quantiles and local load-error gradients; as first author on "Good Pretraining, Bad SFT", the work shows that better-pretraining checkpoints can rank worse after supervised fine-tuning in a 30B MoE pipeline. Bibliometric topics skew toward graph methods — advanced graph neural networks (10 mentions) plus graph theory and algorithms (6) and graph theory and applications (5) — while the current field assignment splits evenly between LLM training systems and LLM training, suggesting an earlier graph-methods base now applied to LLM systems. Evolve's reproducibility assessment is cautious (mean readiness 38 across 2 papers), with both verdicts flagging single-seed, single-scale evidence. Citation-based output and impact scores are not yet computed; h-index 4 and i10-index 2 reflect an early-stage record.
Collaboration pattern
The pattern is a small, stable team: 8 distinct coauthors in the recent window, with repeat collaborators Pit Neitemeier, Philipp Scholl, Jiaze Li, Alessio Serra, Jonas Knupp and Sascha Wirges. Maskey takes first author on one 2026 paper and last author on the other, both multi-author systems work — consistent with a core LMU group around LLM training pipelines. Collaboration scores 44/100, mid-range, reflecting depth over breadth.
Research keywords
Highlights
Exact global quantiles for MoE routing via EQB/LEI
Better pretraining can rank worse after SFT
Single-seed evidence flagged across both studies
Timeline
- 2022Earliest year in the OpenAlex record, with 3 recorded works.
- 2023Lowest annual output in the record: 1 work.
- 2024Output recovers to 3 works, matching the 2022 level.
- 2025Steady year with 3 works, keeping the recent three-year total at 12.
- 2026Highest year on record with 6 works, including both Evolve-reviewed MoE training papers (first- and last-authored).
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 counts, topics and co-author data come from OpenAlex only (identity: openalex); no other citation source is included.
- Output and impact axes are unscored because the citation record is thin (h-index 4, i10-index 2); no citation level is implied.
- Both reviewed papers are 2026 arXiv preprints; peer-review status and final venues are not in the provided data.
- The country field reads 'US' while the affiliation is Ludwig-Maximilians-Universität München; the geographic metadata is unresolved and this profile is name-resolved from public bibliographic data.