Adobe ML researcher: bandit-based LLM routing, skill retrieval for agents, and efficient long-context inference
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
33
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
This axis would measure publication volume and productivity relative to field norms. The normalized data needed to score it is not yet available; only raw aggregates (h-index 33, i10-index 94) exist, so no score is reported.
Impact—
This axis would measure citation influence — total citations, field-normalized percentiles, and reach of the corpus. The snapshot lacks the normalized citation data required for scoring; raw counts (h-index 33, i10-index 94) are shown without a score.
Focus0/100
Focus measures how concentrated output is by field, and this profile sits at 0: the OpenAlex field split is exactly even — 0.5 for machine learning (online learning / bandits for LLM systems) and 0.5 for machine learning — LLM agents and information retrieval. The three Evolve-reviewed 2026 preprints add a third line, efficient long-context sparse-attention inference, confirming a deliberately broad, LLM-infrastructure-oriented spread rather than a single specialized track.
Collaboration58/100
Collaboration captures co-authorship breadth and network position, scored 58 here. The recent window records 202 works with 17 distinct co-authors — a team-based industry pattern. Recurring collaborators include Ryan A. Rossi, Wang Wei, Hoda Eldardiry, Soumyabrata Pal, Koyel Mukherjee, and Branislav Kveton. All three Evolve-reviewed papers list him in middle positions without corresponding-author status, consistent with a contributory role inside larger groups.
Activity100/100
Activity measures recent output intensity and is maxed at 100. Counts by year climb from 36 works in 2020 and 52 in 2021 to 61 in 2022, then 33 in 2023, 51 in 2024, 74 in 2025, and 77 in 2026 — the record's highest. The last three years alone total 202 works, an unusually high sustained rate for a single researcher.
Reproducibility55/100
Reproducibility averages Evolve's readiness assessment across reviewed papers: mean 55 over 2 scored papers, and all three catalog entries are rated medium. The verdicts explain why — diversity-routing gains demonstrated on only 75 queries, bandit proofs with gaps and Nectar Table 1 gains within noise, and sparse-prefilling evidence from single runs with no code released.
Research summary
Franck Dernoncourt is a researcher at Adobe Systems (United States) whose recent work targets machine learning systems for large language models: online learning and bandit algorithms that route prompts to LLM experts, diversity-aware skill retrieval for LLM agents, and sparse-attention prefilling for efficient long-context inference. His OpenAlex profile (identity: openalex; ORCID 0000-0002-1119-1346) shows an h-index of 33 and an i10-index of 94 over a record spanning 2011–2026, with per-year output climbing from 36 works in 2020 to 61 in 2022 and 77 in 2026; the last three years total 202 works, the basis for a maximal activity score. Top-assigned topics are Topic Modeling (295), Natural Language Processing Techniques (162), and Multimodal Machine Learning Applications (137), an NLP-rooted trajectory now aimed at LLM infrastructure. His field split is exactly even (0.5/0.5) between online learning/bandits for LLM systems and LLM agents/information retrieval, and the three Evolve-reviewed 2026 preprints add efficient long-context sparse-attention inference — a breadth consistent with the low focus score. Evolve's reproducibility assessment averages 55 across 2 scored papers, with all three entries rated medium: gains shown on only 75 queries, proof gaps with improvements within noise, and single-run evidence without released code. He publishes in team settings — 17 distinct recent co-authors, with Ryan A. Rossi, Wang Wei, and Branislav Kveton among recurring collaborators — in middle-author, non-corresponding roles on the reviewed papers.
Collaboration pattern
Dernoncourt works in team-based industry settings: the recent window records 202 works with 17 distinct co-authors, and the collaboration dimension scores 58. Recurring collaborators include Ryan A. Rossi, Wang Wei, Hoda Eldardiry, Soumyabrata Pal, Koyel Mukherjee, and Branislav Kveton, among twelve frequently listed names. On all three Evolve-reviewed 2026 papers he appears as a middle author without corresponding status, consistent with a contributory role in larger group efforts.
Research keywords
Highlights
Diversity-aware skill routing for LLM agents
Online learning with LLM experts under limited feedback
CRISP: sparse prefilling for long-context inference
Timeline
- 2011Earliest year in the OpenAlex counts-by-year record, with 1 indexed work.
- 2016Output reaches 13 works, then the record's highest mark before 2020.
- 2020Output accelerates to 36 works, up from 15 in 2019 — the start of a steep climb.
- 202261 works, the highest pre-2025 total in the record.
- 202677 works, the record's peak; includes three Evolve-reviewed arXiv preprints on LLM skill routing, bandit-based LLM routing, and sparse prefilling.
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 derive from an OpenAlex snapshot (identity: openalex) matched via ORCID 0000-0002-1119-1346; name-based disambiguation may still merge or split records.
- Output and impact dimension scores are unavailable; only raw citation aggregates (h-index 33, i10-index 94) are shown, and citation counts typically lag for recent work.
- All three reviewed papers are 2026 arXiv preprints with middle-author, non-corresponding roles; no peer-reviewed venue assessment is included in this snapshot.
- recent3yWorks (202) and per-year counts follow OpenAlex coverage windows and may overcount or undercount near the current date.