Google researcher spanning production recommender systems and LLM-driven black-box optimization
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
8
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: this profile lacks a verified per-paper publication and venue breakdown, so total output, venue quality and authorship roles cannot be aggregated beyond the per-year work counts OpenAlex already provides.
Impact—
Not yet scored: per-paper citation data is unavailable; only snapshot-level h-index 8 and i10-index 8 exist, too thin to measure citation influence, downstream adoption or field-weighted impact.
Focus0/100
The focus score is 0, reflecting a deliberately two-track portfolio rather than concentration: the field distribution splits exactly 50/50 between industrial recommender systems and LLM-agent and black-box optimization work. OpenAlex topics confirm the spread — 22 works in recommender systems and techniques, 22 in information retrieval and search behavior, 11 in bandit algorithms, 10 in topic modeling, plus 2 in AI ethics. That breadth suits industrial research but limits single-field depth.
Collaboration60/100
Collaboration scores 60 on a moderately sized network: 19 distinct coauthors in the OpenAlex snapshot. The frequent-coauthor list is Google-centric — Sanjay Surendranath Girija and Yi Wu each recur twice, alongside Aniruddh Nath, Li Wei, Yanhao Jiang, Shawn Andrews, Aditya Mahajan, Mohit Sharma, Zheng Ren and Zhiyu Hu. Both catalogued 2026 papers are multi-author industrial efforts in which Heldt takes neither first nor corresponding position (middle and last author respectively).
Activity87/100
Activity is the strongest dimension at 87. OpenAlex counts 19 works in the last three years, and the per-year distribution shows a steep recent ramp: 1 work in 2009, sparse output through 2023, then 5 works in 2024, 2 in 2025, and 12 already indexed in 2026. Both Evolve-reviewed papers are 2026 arXiv preprints, consistent with an intensive current production cycle.
Reproducibility38/100
Reproducibility scores 38, the mean readiness rating across the 2 papers Evolve has reviewed. The signals are mixed: the lightweight ranking-heads paper is rated low readiness, with the verdict noting thin offline statistics and claims resting on YouTube production evidence; the code-to-harness paper is rated medium — methodologically careful, though narrow in scope and with production claims that remain externally unverifiable.
Research summary
Lukasz Heldt is a researcher at Google (United States) whose indexed record splits evenly across two tracks: industrial recommender-systems engineering and machine learning for LLM agents and black-box optimization — the OpenAlex field distribution assigns each 0.5. Evolve has reviewed two of his 2026 arXiv papers. As a middle author on 'Lightweight Ranking Heads', he contributed to frozen-backbone adapter heads that reportedly cut YouTube multi-task experiment cycles from 24 to 11 days, though the review notes thin offline statistics. As last author on 'Code-to-Harness', his team distilled LLM optimization practice into a frozen text harness that reduces black-box simple regret by roughly 48%, judged methodologically careful but narrow. OpenAlex topics place him in recommender systems and techniques (22 works), information retrieval and search behavior (22), bandit algorithms (11), topic modeling (10), and a small AI-ethics strand (2). Activity is the standout signal: 19 works in the last three years, including 12 indexed in 2026, for an activity score of 87. The citation footprint remains modest (h-index 8, i10-index 8), the coauthor network counts 19 distinct collaborators, and reproducibility is the clear weakness — a mean readiness score of 38 across the two reviewed papers.
Collaboration pattern
Heldt's recorded collaboration is compact and Google-centered: 19 distinct coauthors in the OpenAlex snapshot, with Sanjay Surendranath Girija and Yi Wu recurring twice each, alongside Aniruddh Nath, Li Wei, Yanhao Jiang, Shawn Andrews, Aditya Mahajan, Mohit Sharma, Zheng Ren and Zhiyu Hu. In both catalogued 2026 papers he is a contributing coauthor — middle author on the recommender-systems paper and last author on the optimizer-distillation paper — never first or corresponding, consistent with team-based industrial research.
Research keywords
Highlights
Lightweight ranking heads compress YouTube multi-task experimentation from 24 to 11 days
Code-to-Harness distills LLM optimizers into a frozen text harness
A two-track industrial footprint in retrieval and optimization
OpenAlex topics show 22 works on recommender systems and techniques and 22 on information retrieval and search behavior, alongside 11 on bandit algorithms and 10 on topic modeling — a footprint pairing production ranking engineering with optimization methodology, plus 2 works on AI ethics.
Timeline
- 2009Earliest work indexed in the OpenAlex profile (1 work recorded this year); the record then stays sparse for years.
- 2024Output picks up: 5 works indexed this year, the first clear step of the recent ramp.
- 20252 works indexed, sustaining momentum between the 2024 rise and the 2026 peak.
- 2026Peak activity: 12 works indexed, including both Evolve-reviewed arXiv papers — lightweight ranking heads for production recommenders and the Code-to-Harness optimizer distillation.
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.
- Bibliometrics are derived from OpenAlex only (identity: openalex), with ORCID 0009-0003-0593-7345; the profile is name-resolved from public bibliographic data and may not capture non-indexed work.
- Only 2 papers have been independently reviewed by Evolve; the reproducibility mean of 38 and all qualitative verdicts rest on this small sample.
- The output and impact dimensions are unscored (dimMissing); per-paper citation counts are unavailable, so h-index 8 and i10-index 8 are the only citation signals shown.
- 2026 counts are current-year and may include early-access or preprint records; the lone 2009 entry and multi-year gaps make long-term career inference unreliable.