Samsung researcher in parameter-efficient fine-tuning: LoRA merging and reasoning-preserving LLM adaptation.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
3
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—
Output scoring is not available yet. This axis would measure total and recent publication counts; the snapshot shows only an OpenAlex h-index of 3 and i10-index of 3, too thin to score a volume-based axis.
Impact—
Impact scoring is unavailable: citation counts and per-paper influence data are missing from the snapshot. This axis would measure citations, citation velocity, and uptake of the parameter-efficient fine-tuning work.
Focus0/100
The focus dimension scores 0 of 100. The field distribution splits evenly — 0.5 for machine learning (LLM fine-tuning) and 0.5 for machine learning (model merging / parameter-efficient fine-tuning) — and OpenAlex topics scatter across five clusters: Domain Adaptation and Few-Shot Learning (8), Advanced Neural Network Applications (3), Multimodal Machine Learning Applications (3), Topic Modeling (2), and Reinforcement Learning in Robotics (2). The two catalog fields are adjacent in practice, but no single field yet dominates the record.
Collaboration39/100
Collaboration scores 39 of 100. The OpenAlex record lists 6 distinct coauthors, and the frequent-collaborator set — Srinivas Chappidi, Wenzhi Fang, Nicholas Tzou, Avinash Amballa, Yashas Malur Saidutta, and Wenbo Li — suggests a compact, single-group pattern. On both reviewed 2026 arXiv papers the author holds a middle position without corresponding status.
Activity57/100
Activity scores 57 of 100, with 6 works in the trailing three years. By-year counts run 1 (2017), 3 (2018), 2 (2023), 1 (2024), 1 (2025), and 4 (2026): a visible 2019–2022 gap, then renewed output that is accelerating in 2026, including both reviewed arXiv papers.
Reproducibility55/100
Reproducibility scores 55 of 100, averaged over the 2 reviewed papers, each rated medium readiness. Evolve flags small-scale, single-seed evidence for the null-basis LoRA method and unreported variance for the budget-aware merging results, so replication looks principled but is not yet demonstrated at scale.
Research summary
Lazar Valkov is a researcher at Samsung (United States) whose recent work centers on parameter-efficient fine-tuning and model merging for large language models. The Evolve catalog includes two 2026 arXiv papers with the author in a middle position: Reasoning-Preserving Fine-Tuning of Post-RL LLMs with Null-Basis LoRA, which freezes a null basis of reasoning activations into LoRA to preserve reasoning gained during RL post-training, and Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks, which introduces a data-free Net Utility metric for rank allocation that beats uniform merging by +2.1/+2.2 on average. Bibliometrics from OpenAlex show an h-index of 3 and i10-index of 3, six distinct coauthors, and by-year counts of 1 (2017), 3 (2018), 2 (2023), 1 (2024), 1 (2025), and 4 (2026) — renewed, currently accelerating output after a 2019–2022 gap. Topic assignments are led by Domain Adaptation and Few-Shot Learning (8), followed by Advanced Neural Network Applications and Multimodal Machine Learning Applications (3 each), Topic Modeling (2), and Reinforcement Learning in Robotics (2). Both reviewed papers carry medium reproducibility readiness and a mean reproducibility score of 55; Evolve flags single-seed evidence and unreported variance as the main limitations.
Collaboration pattern
The collaboration pattern is compact: 6 distinct coauthors in the OpenAlex record, with the same six names — Srinivas Chappidi, Wenzhi Fang, Nicholas Tzou, Avinash Amballa, Yashas Malur Saidutta, and Wenbo Li — recurring as frequent collaborators, consistent with a single research group. On both reviewed 2026 arXiv papers the author sits in a middle position without corresponding status, suggesting contribution within a team-led effort rather than independent leadership.
Research keywords
Highlights
Null-basis LoRA preserves post-RL reasoning through SFT
Data-free Net Utility ranks LoRA merging budgets
Uniformly medium reproducibility across the 2026 pair
Both reviewed papers carry medium reproducibility readiness, giving a mean reproducibility score of 55 over 2 papers: principled methods, but small-scale, single-seed or variance-free evidence so far.
Timeline
- 2017First year with a recorded publication in the OpenAlex by-year counts (1 work).
- 20183 works recorded — the highest annual count in the snapshot before 2023.
- 20232 works recorded, marking the return of output after a 2019–2022 gap in the counts.
- 20241 work recorded; output holds at a single-paper annual pace.
- 20251 work recorded, the lead-in year to the 2026 arXiv pair.
- 20264 works recorded, including both Evolve-reviewed arXiv papers on null-basis LoRA fine-tuning and budget-aware LoRA merging.
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 are computed from OpenAlex (identity: openalex); no ORCID is linked, so this profile is name-resolved from public bibliographic data.
- Output and impact dimensions are unscored because the snapshot lacks sufficient volume and citation data.
- The Evolve review set covers only 2 arXiv papers from 2026; h-index (3) and i10-index (3) reflect an early-stage record.
- By-year counts show a 2019–2022 gap with no recorded works, so career-stage or continuity inferences are unreliable.