Leiden University RL researcher: expert-guided planning and verifiable-reward training for small search agents
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
4
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 snapshot lacks complete publication-volume coverage (e.g. a full works list with authorship positions) needed for this axis. It would measure total research output — number of publications per year and the researcher's role on each.
Impact—
Not yet scored: citation-side coverage is too incomplete for this axis in the current snapshot. It would measure scholarly influence — citations accumulated across the corpus, the h-index trajectory, and uptake by other researchers over time.
Focus0/100
Focus scores 0, the lowest axis on the radar, because the OpenAlex field distribution splits evenly (0.5/0.5) between "machine learning (reinforcement learning)" and "machine learning/NLP (retrieval-augmented agents, reinforcement learning)". Topically, however, the work clusters tightly: reinforcement learning in robotics leads with 12 topic hits and robotic path planning follows with 8, so the 0 reflects labeling across two adjacent sub-fields rather than genuinely scattered interests.
Collaboration36/100
Collaboration scores 36, on the lower end: only 5 distinct co-authors appear across the 9 works from the last three years, indicating consistently small teams. The recurring collaborator list centers on Thomas Moerland, Aske Plaat, Gaurisankar Jayadas and Sandheep P, consistent with the Leiden University affiliation; the cataloged 2026 papers place the researcher first author once and middle author once.
Activity67/100
Activity scores 67. The OpenAlex counts-by-year series shows steady growth: 1 work in 2020 and 2021, 2 in 2022, a peak of 6 in 2023, then 2 in 2024, 4 in 2025 and 3 in 2026, with 9 works in the most recent three-year window. The profile has moved from occasional publications to a sustained yearly cadence without yet reaching high-volume territory.
Reproducibility55/100
Reproducibility averages 55 across the 2 papers Evolve has screened, both rated medium readiness. The RLVR search-agent paper is praised for careful, honest ablations but rests on one model, one dataset and three seeds; GEM-MPC's headline gains depend on heavily overlapping confidence intervals from 5 seeds. Methods and results appear reproducible in principle, yet statistical power is the recurring weakness.
Research summary
Álvaro Serra-Gómez is an early-career researcher at Leiden University (NL) working on reinforcement learning, with an OpenAlex profile spanning 2020–2026 (h-index 4, i10-index 4; annual works rising from 1 in 2020 to 4 in 2025 and 3 in 2026). The record splits evenly (0.5/0.5) between reinforcement learning in robotics and retrieval-augmented, RL-driven language agents, and topic counts cluster on RL for robotics (12) and robotic path planning (8), with smaller clusters on multi-agent control, doctoral education, and health research. Two 2026 arXiv papers are reviewed in the Evolve catalog: a first-author study of expert-guided MPPI planning with value-gated distillation (GEM-MPC), and a middle-author study applying GRPO with verifiable rewards to a Qwen3.5-0.8B search agent, lifting exact-match from 0.092 to 0.352. Evolve's reproducibility screening averages 55 across two papers (both rated medium), with reviewers crediting careful ablations while flagging small seed counts and overlapping confidence intervals. Collaboration is compact — five distinct co-authors in the last three years — and the radar scores activity at 67 and reproducibility at 55, while output and impact axes remain unscored for lack of coverage.
Collaboration pattern
Collaboration is compact: only 5 distinct co-authors appear across the 9 works of the last three years, giving a low radar collaboration score of 36. Recurring names include Thomas Moerland, Aske Plaat, Gaurisankar Jayadas and Sandheep P, consistent with the Leiden University affiliation. Authorship in the catalog is mixed — first author on the GEM-MPC method paper, middle author on the verifiable-reward search-agent study — suggesting a hands-on experimental role within small teams rather than large consortium work.
Research keywords
Highlights
GEM-MPC: expert-guided MPPI planning with value-gated distillation (first author)
Verifiable-reward RL for small search agents (middle author)
Timeline
- 2026First-authors GEM-MPC (arXiv), combining expert-mixture MPPI planning with value-gated distillation of stale planning targets.
- 2026Co-authors (middle position) the RLVR study on small search agents, lifting Qwen3.5-0.8B exact-match from 0.092 to 0.352.
- 2026Evolve screens both cataloged 2026 papers, rating each medium for reproducibility readiness with a mean score of 55.
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 indicators (h-index 4, i10-index 4, counts-by-year, topics, field distribution) come from an OpenAlex identity resolution only; the output and impact axes are unscored because the underlying coverage is insufficient.
- The frequent-coauthor list includes an unaccented variant of the researcher's own name ("Alvaro Serra-Gomez"), a likely name-disambiguation artifact; this profile is name-resolved from public bibliographic data and only 2 papers carry full Evolve reviews.
- Both cataloged papers are 2026 arXiv preprints; peer-reviewed publication status and long-term citation impact are not yet visible in this snapshot.
- The reproducibility score of 55 reflects Evolve reviewer judgments on the 2 screened papers only, not an independent replication attempt.